11 Commits

Author	SHA1	Message	Date
Ben Duncan	8ab96b29a1	fix: replicate every JetStream KV bucket across the cluster (#1117) ... * fix: replicate every JetStream KV bucket across the cluster A KV bucket created without a replica count gets one replica, and one replica lives on one JetStream-chosen server recorded in no config. Losing, restarting or partitioning that server makes the bucket return "no responders available" from every node at once, so a single-node event becomes a cluster-wide outage. The reconciler leader-lease buckets were the worst case: acquiring a lease is a write, so a lease bucket without a quorum cannot be acquired by anyone, and every reconciler sharing it stops cluster-wide at exactly the moment a node has failed. The cluster's node count is now declared once, in config.LoadConfig, and read by the one helper every bucket creation goes through. Every process that creates a bucket loads the cluster config first, so no service can forget to pass it and no signature has to carry it. An undeclared size is an error at creation rather than a silent fall back to one replica. Existing buckets are raised rather than left behind: a service opening one raises it, the daemon sweeps them all after formation, and "spx admin kv replicas --repair" does it on demand. Raising reads the live stream config and changes only the replica count, and never lowers one. Removes the ~30 replicas/clusterSize parameters that only forwarded a number, and JetStreamManager.UpdateReplicas, which set len(nodes) with no clamp and would have failed outright on a cluster of six or more. Proved by a repo-wide invariant test pinning bucket creation to the one helper, a whole-cluster sweep in every multi-node nightly cell, and a new kvquorum suite (cell 33) that takes away the node leading the most buckets and requires the survivors to keep answering for all of them. * chore: name cell 33 in the nightly dispatch input * fix: a KV bucket that cannot be raised must still open The first version of the replica fix treated a raise it could not place as a hard error, and on a real three-node cluster that turned the fix into the outage it prevents. A node whose config named a fourth server wanted four replicas, NATS answered "no suitable peers for placement" for three, and the daemon retried for its full five-minute budget without finishing bootstrap. The bucket was present, quorate and serving on three nodes throughout. A config naming a node before that node serves is the ordinary middle of growing a cluster, so that is a state to tolerate rather than refuse. The line is now drawn between the two cases instead of around both. Creating a bucket below the cluster's replica count still fails, because that is the defect and a fallback there reintroduces it where it is hardest to notice. Failing to raise an existing bucket warns and hands the bucket back, and the raise is retried on every open besides the daemon sweep and --repair. This matters most in kvlease, which runs on every reconcile tick: refusing there would stop every reconciler sharing a lease bucket that was working. getOrCreateBucket now attaches before it creates. A create asking for more replicas than the cluster can place is refused before the server looks at the name, so with create-first "this bucket already exists" was an inference from which error came back, and the wrong one comes back on a single-node server. RaiseAllBucketReplicas attempts every bucket and joins the failures, and --repair keeps going too. Stopping at the first would leave the rest of a cluster short because one bucket could not be placed. Also adds nats.ServerName/NodeFromServerName beside the template that renders server_name, with a test tying them together. The kvquorum suite had mapped stream peers to nodes by comparing them to node names, which they are not: they are the node name behind a "spinifex-nats-" prefix, so nothing matched and the node-loss test could not choose a victim. * test: wait for the right NATS peer count after a node rejoins /routez reports a node's routes to its peers, not its own membership, so a fully formed three-node cluster is two peers per node. The kvquorum suite waited for three and spent its whole five-minute rejoin budget on a cluster that had already reformed. Every other caller passes len(Nodes)-1; the helper's doc comment now says so. The assertion the test exists for had already passed: node1 was leading 18 of 47 buckets when it was taken away, and node2 and node3 each still answered for all 47. * fix: replicate KV buckets across the largest odd count a cluster can place A bucket was replicated across min(nodes, 5) servers, which stores a fourth full copy on a four-node cluster and still survives exactly one loss — the same tolerance R3 gives for three quarters of the write amplification. A Raft majority needs an odd count, so the rule becomes the largest odd count the cluster can place, capped at five: 1, 3, 3, 5, 5. It is one expression with two entry points, so no caller can hold a stale copy of the policy and keep passing after the policy changes. Three retry loops could not tell a configuration fault from a cluster that had yet to form. An undeclared cluster size never becomes declared by waiting, so each of them spent its whole budget on one and then reported quorum as the cause. They now stop on a permanent error and say what is actually wrong. This is what the integration suite was hitting: it builds its gateway by hand and never loads a cluster config, so the size was never declared and the IAM retry sat in its sleep until the test timed out. Both ends are fixed. The audit reported configuration alone. A bucket at three replicas with one peer offline and one lagging is one failure from unavailable and read as ok. BucketReport now carries the leader and a count of the peers JetStream itself calls current and reachable, quorum is measured against that count, and spx admin kv replicas exits 0, 1 or 2 so a deployment gate can tell a problem --repair fixes from one needing a human. Declaration is write-once: daemon.Start refuses a conflicting redeclaration because it is about to create buckets, and it declares before JetStream init rather than after it. config.LoadConfig warns instead, because it also reads config files that are not this process's own cluster. spx admin init --nodes 2 is refused — no replica count on two nodes both survives a loss and accepts writes — and single-node replication is asserted in the single e2e suite, so it runs in every nightly host permutation. * fix: withhold the vLLM serving image until its artifact is published The catalog offers ubuntu-26.04-vllm-serving-x86_64, and the qcow2 it names 404s on iso.mulgadc.com, so spx admin images import would fail for every operator who tried it. Its GPU siblings are published; this one never was. The entry moves to WithheldImages rather than being deleted, because the entry is the definition — the URL, the checksum and the two tags resolveServingAMI filters on. Publishing the artifact is then a move back, not a retype. Bedrock already handles the image not being registered: resolveServingAMI returns ErrServingAMINotFound, which is the honest answer while there is nothing to launch. The bedrock tag contract is still pinned, now against the map the definition lives in, and it additionally asserts the entry is not offered for import. The e2e reachability walk covers AvailableImages alone, so a withheld entry is out of it by construction. * fix: drop the cluster size from the call sites dev added Two constructors on dev still pass a replica count that no longer exists as a parameter: the cluster size is a declared process fact, so EKSServiceDeps has no ClusterSize field and NewIAMServiceImpl takes three arguments. The merge is textually clean and breaks the build, which is the shape of collision a signature removal always has. * docs: explain what an operator sees when a cluster server fails The replication section said how many copies a bucket has and stopped there, which leaves the question an operator actually has during an incident: one server is gone, what still works. It now covers the majority side carrying on, an isolated server refusing changes on purpose and why that is the behaviour you want, guests continuing to run while anything that has to be recorded does not, the risk of reading state from a server that is cut off, and a returning server catching up by itself. The surrounding prose reads from the operator's side rather than the implementation's: what you see and what to do, not which consensus protocol is underneath.	2026-09-28 12:34:55 -05:00
Ben Duncan	021f439bab	feat: recover instances automatically from a node that stopped heartbeating (#1111) ... * feat: recover instances automatically from a node that stopped heartbeating A node that dies hard takes its guests with it, and nothing brings them back until an operator brings that node back. Every daemon now runs the same small reconciler: it reads the instance records, selects desired-running instances whose owner is stale, and relaunches them here. The claim is a CAS on Status.LastNode and that is the whole of the coordination. Several survivors race one instance, exactly one revision wins, and the losers find a record naming a live owner and stop considering it. A claimant that dies mid-launch leaves the record naming a node that is itself stale, which is the same candidate the next pass derives again, so there is no lease to expire and nothing to reconcile by hand. Ownership is not permission to write. The volume leases still decide whether the guest can move, and VolumeMounter.Mount runs before startQEMU, so an attempt that loses the volume race has touched no tap, no OVS port and no OVN binding. A storage failure is not a host failure and wants the opposite response. A guest paused by werror=stop publishes Health.IOErrorSince, and a record carrying it is refused: the volumes are the same objects in the same store from anywhere, so relaunching moves the pause rather than ending it. A node whose own object store is not answering stands down entirely, which separates a store broken here from a store broken everywhere without either having to be diagnosed. A launch that does not complete backs this node off that instance, doubling to a cap, because nothing on the mount path can yet say which of the two it hit. Recovery is off by default and inert below three nodes. It also waits until it has seen its peers live, or five minutes, so the first node back from a coordinated shutdown does not recover guests from nodes that are still booting. * test: prove instance recovery moves a host failure and refuses a storage failure The e2e half of automatic instance recovery, asserted from both sides because a reconciler that cannot tell the two failures apart passes one and fails the other. InstanceAutoRecovery takes a node down and requires the guest back on exactly one survivor, identified by the qemu process rather than by what the API reports. InstanceRecoveryRefusesStorageFault stops predastore cluster-wide first, so the guest is paused by werror=stop when its node goes away, and requires that nothing moves. Also fixes the settle window, which re-evaluated every pass and so held a daemon back for its first five minutes over exactly the failure the reconciler exists to handle. It now latches once every peer has been seen live. harness.InstanceHostingNodes replaces the single-node lookup so a caller can assert there is exactly one host rather than assume it, and skip a node it has deliberately taken down. * fix: make automatic instance recovery a cluster-wide setting Recovery was read from the per-node Config, so it had to be written three times on a three-node cluster, once inside each [nodes.<name>] table. A top-level [recovery] table — the obvious thing to write, and what the e2e suite writes — was parsed into nothing and the daemon reported recovery disabled on every node with the setting plainly in the file. Cluster-wide is also the honest shape. A node with recovery off is not opting out of anything: recovery is what survivors do to a dead node's instances, so the setting belongs where there is one of it. * test: drive real I/O into the storage-fault case, and freeze the store rather than stop it Measured on a three-node cluster: predastore stopped cluster-wide for twelve minutes against idle guests paused none of them. Viperblock serves what it has cached and an idle guest issues nothing that reaches the object store, so the test asserted that nothing moved in a run where nothing was ever faulted. The victim now launches with cloud-init user-data that writes to its root volume continuously. User-data rather than SSH because a routed-NAT cell gives guests no public address, so a test that had to log in could not run on half of them. The fault is SIGSTOP rather than a unit stop, for the reason the storagefault suite already gives: a frozen process holds its connections open and answers nothing. Under that fault and that load the first BLOCK_IO_ERROR arrived 4m47s later, so the pause budget is now ten minutes rather than three. * test: give instance recovery its own suite and its own nightly cell Both tests take a node away and one of them freezes predastore cluster-wide, so anything sharing the environment reports that outage as its own failure — the same argument storagefault already makes for being its own package. They are also slow by nature: a dead owner has to be seen stale twice, a guest has to boot, and the storage case waits for a pause measured at 4m47s and then holds four minutes to prove nothing moved. Roughly twenty minutes on top of a permutation cell budgeted at thirty-five. Cell 31 is three-node and static, matching storagefault's, with a 75-minute job budget over a 60-minute suite timeout. The suite is eligible on multi-node environments and deliberately not in the nightly permutation set. * fix: give a guest long enough to open a volume the object store is slow to serve QEMU opens every drive before its monitor answers a greeting, so the QMP greeting deadline is really a deadline on opening the guest's volumes. At 30s it was shorter than a cold viperblock open costs on a cluster that has lost a node, where every object read is served degraded: measured at 35s with one node of three away. The launch killed the guest five seconds short of ready and then did it again on every retry, which is a recovery that can never complete. A network-backed drive now gets its own deadline for the same reason VFIO already had one — synchronous work before the monitor answers — and the instancerecovery suite is what found it. Also make the suite's predastore thaw best effort per node. It runs after a node has been taken away, where there is no process to signal, and a node it cannot reach must not stop it restoring the rest. * docs: name the two newest cells in the dispatch input The cells input is the operator's index of what a number means, and it stopped at 29 while 30 and 31 were added. * test: prove the guest is writing before freezing the store in the recovery suite The storage-fault test froze the object store as soon as RunInstances reported running, which is when QEMU has launched rather than when the guest has booted. On the nightly cluster the freeze landed while the guest was still in UEFI, so the only drive it could reach was the EFI varstore — and pflash has no werror=stop to pause on, so the guest never paused and the test reported a product failure it had caused itself. The load driver now prints to the serial console after each pass it has written and synced, and the test waits for that before freezing. The console is read through GetConsoleOutput, so this still needs no login and still runs on cells whose guests have no public address. * feat: recover instances automatically with nothing to configure The switch is gone. An instance whose host has failed is down whether or not recovery is enabled, so the only outcome a disabled reconciler can produce is leaving it down — which is the state recovery exists to end, not a safer one to fall back to. Against that, the switch cost a cluster that can be configured into never recovering anything, with no symptom until a node dies. This feature already hit that once: a top-level table parsed into nothing and every node reported recovery off with the setting plainly in the file. What makes recovery inert is now only facts about the cluster — no cluster config, no JetStream, or fewer than three nodes — and each says in the log which one applied. The e2e no longer writes config or restarts daemons: it waits for every node to report the reconciler active on a cluster nothing configured, so reintroducing a switch fails there as well as in the unit tests. * feat: stop retrying a recovery that is never going to work Every failed attempt used to go to the same exponential backoff and stay there. An instance nothing could launch was retried every fifteen minutes forever, still observed-running on a node that is gone, with no account of it anywhere the customer or an operator would look. A failure is now classified and counted. Two classes are read from what the launch actually reports: the node it was taken from still holds a volume lease, which clears on that lease's clock and is retried on a faster one, and this node has no room, which another survivor may. There is no terminal class, because nothing on this path can prove one — the launch collapses everything else into ServerInternal, so a verdict of "never going to work" would be a guess, and the guess that abandons a recoverable instance is the expensive one. The budget bounds the unknown case instead. Every class counts against it, since a class that did not would be the old unbounded retry under a narrower condition. Exhausting it leaves the instance where AWS leaves one whose host failed and could not be brought back: stopped, carrying a StateReason that names the cause and the last error, one StartInstances from running again. Writing the desired state down is also what makes one node's decision cluster-wide, because selection reads desired state — the same property the claim already has. A record another node has since claimed is never touched, so a node exhausting its attempts cannot stop a guest that is now running elsewhere. * fix: give a guest the time for every cost it pays before its monitor answers A GPU instance booting from a network-backed root volume pays both costs: QEMU opens every drive across the object store and then pins the whole of guest RAM, one after the other, all before the monitor answers. The deadline took whichever was larger, so such a guest was given less time than it needs. That the two floors are both 180s is why nothing has been seen yet — the VFIO branch has always returned at least the network-drive deadline. It is the same defect either way, and the symptom when it lands is a GPU instance SIGKILLed on every attempt while its plain sibling launches, which reads as a GPU fault rather than as a deadline. The costs are now added, the plain QEMU startup is counted once rather than replaced, and the cap covers both floors together so it cannot clip a guest before either cost has been waited out. * test: assert the AWS-visible arc of a host failure end to end A customer cannot see which host an instance is on, and AWS never tells them. What they can see is DescribeInstanceStatus, so that is where a recovery has to be provable, and the suite never looked at it. Three frames are now required, in order: running with system-status ok before anything is broken, system-status impaired while the owning node is gone, and ok again once a survivor is running it. The middle frame is the one worth having — the window opens when the owner's heartbeat goes stale, which is strictly before any survivor may claim, and closes when the relaunch registers, so it is bounded rather than lucky. Reporting ok throughout would be the more comfortable lie. The availability zone is compared across the first and last frames as well. A recovery that crossed one would break every placement promise made to the customer, and nothing else in the suite would have noticed. * test: prove a half-healthy node fences itself and is never a second writer The two existing recovery cases take a node away by stopping its services, which is the failure this feature handles least interestingly: a node that is gone has no opinion about who owns its volumes. The one that matters is the node that is still up, still running QEMU, still answering its own local clients, and simply cannot reach consensus. That is the state in which two nodes can each believe they own a volume, and the only thing between that and a corrupted disk is that the partitioned node gives up first, on a clock it can evaluate without reaching anything. So this drops NATS on 4222 and 4248 between one node and its peers and nothing else, then asserts the ordering the safety argument rests on rather than the outcome it produces. A watch samples every node every three seconds for the whole run, including the heal, and fails on any single sample naming two hosts. Two guests writing one volume through one object store is data loss already under way, not a race that resolves, and no point-in-time assertion can see one that appeared between its own samples. The injector takes its own nft table rather than touching inet spinifex_filter, because a test that flushed the node's real firewall would leave it either unfirewalled or misfirewalled with no sign of which, and a separate table at priority -300 heals in one atomic delete. A transient systemd timer removes it independently, so a cancelled workflow does not leave a node cut off from its cluster until somebody notices. It proves SSH still works and proves a peer port is genuinely unreachable before anything is asserted, since a rule set that installed cleanly and dropped nothing would read as the system tolerating a partition. The suite budget goes to 90m and cell 31 to 105m to fit the third case. * fix: move an instance's boot state, volumes, logical port and address with it A record-driven launch assumes the node reading the record is the node that wrote it. On a daemon restart that holds. On a recovery onto a survivor it does not, and five things follow from it — one that stops a moved instance booting at all, and four that stop it running anywhere. A load balancer never comes back from a host failure. Its boot configuration is a set of files under the runtime directory and the record holds their paths, so the survivor restores paths that exist only on the node that died. The hook that rebuilds them was reachable from Restore alone, and the cross-node path goes straight to vmMgr.Run — so QEMU is handed a netcfg, an agent env and a CA cert that are not there. The gateway URL baked into that env names the creating node's management bridge as well, which the rebuild corrects because it rebuilds from whichever node is running it. Firing the hook on both paths fixes both, and a rebuild that fails now refuses the launch: a guest that boots without its network configuration is worse than one that does not boot, because the records would say it recovered. The export outlives the guest and holds the volume lease. A returning node is healthy, so it renews that lease indefinitely and the new owner never acquires: the guest is stopped here and refused there, recovery spends its attempt budget on a volume that was never going to be free, and gives up on an instance that was never unrecoverable. So the export is given up too, over a new node-addressed request that tears it down exactly as the fence does. Deliberately not an unmount — an unmount seals, this node's block map is the stale one, and sealing publishes it over the winner's. Unlike the fence it deletes the lease key rather than leaving it to expire, because the lease is still this node's to hand back. The tap outlives the guest, and the tap is what claims the instance's OVN logical port. The binding names one chassis, so two nodes offering the same iface-id do not split the traffic — they contend, and the address answers from whichever won last. It also blocks the layer above: the external IP's host ingress is pruned on the evidence that the owning guest sits on another chassis, and that stays false while this node is still claiming the port. The management address outlives the guest as a cluster-wide reservation, and the release was unconditional. br-mgmt is one flat L2 segment, so the node an instance left was freeing an address still in use and offering it to the next instance anywhere. Only the node an entry names may release it now, with the takeover recorded on arrival so the node actually running the instance can still give the address back. An entry written before the node field existed names nobody and stays releasable, because there is no other holder to collide with. The fence path already cleaned up the tap, so the release fix repairs that path too. * test: prove a load balancer keeps its name, its address and its traffic when its host dies Takes away the node running an internet-facing ALB and asserts the four things a customer of one actually holds: the name keeps resolving, the address behind it does not move, requests through the name are answered again without anybody being called, and the backends are healthy again from the recovered load balancer's point of view. The backends are placed around the load balancer rather than beside it. A load balancer lands wherever a node answers its queue group, so its node is discovered and the backends are relaunched until both are elsewhere; one on the node about to be taken away would be recovered too, and the run could not then tell an ALB that came back serving from one that came back with nothing to serve. DNS is asked of a node's own northstar over the wire, and the HTTP probes resolve through the same server, so the claim under test is what this cluster answers for a name rather than what the machine running the tests was configured to forward to. A watch samples every surviving node every five seconds for the whole run, because an answer that went missing or pointed elsewhere for one reconcile pass is the outage, and no single sample reaches it. The returning node then has to give up what it no longer owns: no second copy of the guest, and no tap still claiming the load balancer's logical port, since OVN binds a port to one chassis and a second claimant contends for it rather than splitting the traffic. The last assertion is that the name still serves with that node back in the cluster. * docs: name the new cell in the nightly's cell list * test: stop the disk-load guest from filling the console it is judged by The load generator announced every write, once a second, for as long as the test ran. The console is a ring buffer this suite reads back, so an hour of that pushes out everything explaining a failure, including the kernel's own account of the I/O error the test exists to produce. Announce twenty times, then keep writing quietly. * docs: say what the OCI guide has been proven on, and what a host failure does The guide read as a support statement for OCI rather than as an account of one tested path, promised seven steps and then gave eight, and said nothing about what happens to a guest or its public address when the node under it fails. State the evidence, state that the nightly guards the datapath and not the cloud, and give the failure behaviour as a table with the reason codes a customer will actually see. Also say plainly what the volume lease is: one running guest and a bounded, logged window, not a write barrier the object store enforces. Someone sizing a workload against this should read that before they read the table. * docs: say why the mount classifier refuses to retry a lease another node holds It reads as an omission — a lease held elsewhere is the most transient refusal there is. It is not. This classifier drives the restore path, where the holder is a node that is running the guest, so retrying would relaunch against a live owner. Recovery wants the opposite and reads the refusal itself. * feat: claim a recovered guest's public address in line, and verify it arrived A recovered instance was reported running before anything moved its public address off the node that died, so for up to fifteen seconds the records said recovered and the address answered nothing. The recovery path now runs the affinity pass itself, which is idempotent and is the same pass the ticker runs. Two things the pass could not do before. It abandoned every remaining address on the first API failure, though a throttle or a denial is per-object and the guests behind the others are just as dark. And it trusted the move's own response, which is OCI agreeing to the request rather than OCI having done it, so an accepted move that never landed read as success. It now re-reads what OCI delivers here and reports the difference. An address that will not arrive makes the guest a system-status failure with the time it went dark, the same shape AWS uses for a host-level fault. The guest passes every check it can be asked about itself, which is exactly why reporting it healthy was the worst available answer. * feat: give a recovery attempt one trace across all three of its phases The volumes opening, QEMU starting and its monitor answering already trace themselves, but nothing tied them to the attempt that asked for them: three orphan traces per recovery, no way to ask which phase a slow one spent its time in, and nothing naming the node the guest came from. A span over the attempt makes them its children, and the reconciler's own log lines carry its trace ID because the handler stamps it from the context. A failed attempt records the fault and the attempt count on the span, so the budget running out is readable as a sequence rather than as twelve unrelated warnings. * fix: finish a fence once its writer finally exits A kill that did not take effect was logged, counted and dropped. The export stayed registered and the lease stayed held, which is correct while the writer lives and wrong forever after it dies: a lease a healthy node keeps renewing never expires, so the volume could not start anywhere. The failure is now re-reported every 30 s until it resolves, and the teardown completes as soon as the writer is confirmed gone. Only a delivered kill is waited on — a refused one says nothing about what became of the writer. * test: stop judging a detached stop or terminate on a transient state Both tests read the instance state immediately after the ack, but the transition runs on its own goroutine, so the read races it in both directions: too early sees running, too late sees the final state. The race gate caught the first of those on a loaded runner. * fix: replicate the volume lease and dirty buckets across the cluster Both were created with the default single replica, so both lived on one node. That node holds the answer to "who may write this volume" and the marker every mount reads, and losing it takes volume leasing and mounting down cluster-wide. A partitioned node was fenced correctly and then no survivor could relaunch its guest: every attempt failed on "read dirty mark: no responders available", for ten minutes, until the budget ran out. The replica count is now the cluster's node count, as the gateway's own buckets already do. Existing buckets are raised in place rather than requiring a wipe. The e2e diagnostics that should have said this are fixed too. The fence and the surrender are logged by the storage service, not the daemon, so the assertions read the wrong journal; and Eventuallyf evaluates its message at the call site, so a ten-minute wait was explained by a log that stopped before it began.	2026-09-27 15:56:09 -05:00
Ben Duncan	ea3e01bb20	fix: reboot a guest by asking it, not by pulling its reset line (#1028) ... * fix: reboot a guest by asking it, not by pulling its reset line RebootInstances was QMP system_reset alone. The guest kernel is never told, so it does not sync, does not unmount and loses every dirty page - which is how a reboot issued seconds after a write loses that write, and how a guest rebooted during cloud-init can fail to come back at all. Guests now launch with -action shutdown=pause, so a clean guest poweroff leaves QEMU paused rather than gone. Reboot presses the power button, waits for the guest to reach that state, then resets and resumes the same process: no volume, tap or state churn, and no window where the guest is missing. A guest that will not go quietly is hard-reset, which is what AWS documents. The button is pressed again every 10s while the guest ignores it. One signal is not a shutdown request on a guest whose acpid has not started, which is the same boot race that produced the failure. Two consequences handled here: a guest that powers itself off now leaves QEMU paused, so the heartbeat quits it; and a guest launched before the flag exits on powerdown rather than pausing, so it is reset without being asked. * fix: answer a reboot when it is accepted, not when the guest is back Asking a guest to shut down takes as long as the guest takes. RebootInstances gives the daemon five seconds to reply, so a reboot that waited for the guest reported a reboot that had in fact succeeded as InvalidInstanceID.NotFound. Manager.Reboot now admits the reboot — the instance exists and is running — and runs the powerdown and reset in the background, which is the contract EC2 documents. The two admission failures are still reported; nothing after them can be, so a reset that fails is logged and left to the heartbeat. The gateway also stops calling a NATS timeout NotFound. Only the absence of a subscriber says anything about where the instance is, which is how the sibling start and terminate paths already read it. * test: prove the churn reboot by its boot ID, not by how fast it happened The subtest waited ten seconds and then read /proc/uptime. A reboot is asynchronous, so the guest answering at that point may be the one that has not shut down yet, and its uptime is whatever it was before — which the test reads as "reboot may not have occurred". Waiting for the boot ID to change is positive proof the guest rebooted and does not depend on how long the guest takes to go down. The uptime assertion stays, now measured against a boot that is known to be new. * test: prove the graceful reboot against real guests and a real engine Two tests the reboot design asked for and the unit tests cannot give. Cell 18 launches two guests that differ only in how they answer the power button. One installs a shutdown hook writing an fsynced line a second, so the file that survives measures how much of its shutdown it was allowed to run — a guest reset instead of asked leaves none of it. The other ignores the power key entirely and must be reset once its budget expires, with the daemon saying so. Each variant asserts the evidence only it can produce, and the slow one asserts the absence of the line the wedged one requires. Cell 27 loads a PostgreSQL instance past its shared buffers with pgbench and reboots it dirty. The probe is an UNLOGGED table: PostgreSQL writes no WAL for one, so a clean shutdown checkpoints it and crash recovery truncates it unconditionally. Rows committed normally survive a hard reset too, so the obvious assertion would pass against the bug. The rds suite and cell 27 both gain budget for the load phase. * fix: stop the process-group test misreading a reaped child as alive The test polled until the child was gone, then checked once more. A task being reaped shows state X for an instant, which processRunning counted as running, so that second look failed about 1 run in 50 under -race. * fix: wait for the recovery retry loop before removing a test's data dir restoreInstances starts the retry loop untracked, and it rewrites the local state file when it ends. A write landing during t.TempDir cleanup failed TestRestoreInstances_* with RemoveAll: directory not empty.	2026-09-14 15:39:27 -05:00
Ben Duncan	18853cef31	fix: close the four defects keeping the e2e nightly red (#1007) ... * fix: stop the e2e analyser inventing failures and let a detach outlast its seal go-junit-report emits <error message="No test result found"> for a === RUN with no matching result line, and attaches the test's own captured log as the body. The analyser counted those as failures, so the baremetal cell reported 25 failures against a junit file recording 1, with log output reading as assertions. Separate the marker from a real error in both the parse and the roll-up, count the unresolved cases and say so in the summary. The gateway gave a detach 30s while the daemon budgets 180s for the ebs.unmount seal and 180s more for the QMP unplug chain, so any detach whose seal replayed a WAL was reported as ServerInternal while it went on to succeed. Also switch four nightly job gates from always() to !cancelled(): always() overrides cancellation, so a cancelled run kept starting jobs. * fix: serialise the e2e workflows on one group and retry a dropped libvirt e2e.yml had no concurrency block at all, so every push to dev could provision alongside every other push and alongside the nightly. All four workflows that build on these hypervisors now share e2e-hypervisors, queued rather than cancelled. A per-environment group is not expressible: GitHub picks the group before it picks a runner, and the environment is a property of the runner. tofu apply also failed outright on a transient libvirt connect, which fails a 45-minute job in 90 seconds and reports it as a test failure. retry_run retries only when that attempt's own output names a connection failure, so a real failure still fails once. Also collect the unfiltered journal from the matrix cells and the node journal from cell-21: the unit filter drops the kernel and OOM lines that explain why one node behaved unlike its siblings, and a passing run's journal is what makes a failing one readable. * fix: build the baremetal junit from the unfiltered suite log run-baremetal-e2e.sh streams the full suite output to suite-output.log and a grep-filtered summary to its stdout, which the workflow tees into test-baremetal.log. That filter anchors --- PASS at column 0 and Go indents every subtest result by four spaces, so the junit was generated from a log with no subtest results in it: go-junit-report saw each === RUN with no outcome and emitted <error message="No test result found">. 17 of cell-28's 39 testcases and 93 of cell-20's 120 were that. Regenerating cell-28 from suite-output.log gives 39 tests, 0 failures, 0 errors. * fix: converge released EIPs instead of trusting a fire-and-forget teardown A disassociate leaves the ENI running, so the orphan sweep's owning-port test never fires for it. The teardown then rests entirely on vpc.delete-nat, which is core NATS and is lost whenever vpcd is restarting — and the intent KV watch is UpdatesOnly, so the same window hides the write that would have driven a repair pass. The dnat_and_snat row and its host /32 route to a released public address survived until something else happened to touch them, which for a live ENI was nothing at all. Give the sweep the second test it was missing: AddEIP is the only writer of dnat_and_snat, so the floating IPs intent asks for are the complete set of addresses that may hold one, and a row for any other address is stale however alive its port. Ports and external IPs travel together in LiveEIPs, both unioned from the start-of-pass snapshot and the fresh re-read, so a mid-pass launch is still spared. The host unbind now runs even when the row had already gone, which is the state a crash between the OVN delete and the unbind leaves. Run that sweep on the startup pass too. Startup is the pass that follows the blind window, so skipping it deferred repair of exactly the changes that were missed. The port-group and ENI-port sweeps stay gated there for the in-flight create they were written for; the EIP sweep re-reads intent immediately before deciding and abandons the whole sweep if that read fails, which is the same protection by a stronger route. * fix: keep a release verify off the hypervisors an e2e run is using release.yml is the fifth workflow that provisions on the ci-single pool and had no concurrency block at all, so a tag push could land a tofu apply on the same libvirt host a nightly cell was mid-provision on. The group goes on the verify job rather than the workflow: the build jobs touch no hypervisor and should not queue behind a nightly. * fix: stop cell-27 handing a 50-minute suite a 50-minute job The rds suite is budgeted 50m in suite-sets.sh and cell-27 took the matrix default of 50 minutes for the whole job, so provisioning, install, the suite and teardown had to fit inside the suite's allowance alone. It never could: the last measured run spent 6 minutes reaching the suite and was killed 44 minutes into it, mid-TestAutomatedBackups. A job timeout reports "The operation was canceled" and nothing else, which is why this cell has no verdict in any nightly rather than a failure anyone could read. cell-27 gets 75 minutes, and the select job now refuses to start a run whose job budget does not clear its suite budget plus a setup allowance. The inner timeout firing first is the property that matters: go test says where it stopped and what failed, and the job timeout says nothing at all. cell-29 already had the same fix applied by hand; nothing stopped the next suite from repeating it. * fix: converge host EIP plumbing against intent, not against the NAT row The released-EIP sweep landed and did not fire. Its premise was that a stranded dnat_and_snat row is what a stranded host route can be found by, and on the path that actually strands one there is no row left. AddEIP scrubs the predecessor row sharing the private IP before installing the new one, so when reconcile re-added the instance's auto-assigned address it silently deleted the released EIP's row — and applyEIPs runs ahead of the sweep, which then walked a table with nothing stale in it while the /32 route and its proxy-ARP stayed up. Give the sweep the host's own inventory. ListEIPIngress reads the EIPs back from the FORWARD accepts EnsureEIPIngress stamps with its own comment, so a binding is found by what installed it rather than by a row some other path is free to delete. Anything plumbed that intent no longer asks for is torn down, which also covers a crash between the OVN delete and the unbind. A failed inventory fails the pass: an empty list and an unreadable one are indistinguishable, and reading the second as the first is a sweep that reports success having done nothing. Close the window inside the pass too. The predecessor scrub now reads the row it is about to drop and releases that address's host state, so the strand does not exist between applyEIPs and the sweep rather than merely being repaired a step later. * fix: retry the release verify's libvirt calls like every other job does The verify job drives the same libvirt provider on the same hypervisors as the e2e cells, where a dropped connection under concurrent load is the steady state and is retried. Here it was a hard failure that reads as a broken release. It was skipped earlier only because the job checks out mulga and not spinifex, so the shared helpers were not on disk — a sparse checkout of .github/scripts costs nothing and makes apply and destroy behave like their e2e counterparts. * fix: do not report a rebooting DB instance available on the pre-reboot agent's health RegisterDBInstance stamped a fresh LastSeen while leaving Agent.EngineHealth at the value the previous agent process reported. After a reboot the new agent registers before it has probed anything, so the record read as a healthy engine with a beat newer than the transition, and the reconciler moved the instance to available while PostgreSQL was still starting up. A client that connects the moment the status turns green then gets "the database system is starting up". A register now discards the previous process's verdict and records when this agent process started. The restart reconciler requires that registration to be newer than the transition, which also closes the window where a beat lands between the engine being stopped and the VM going down. * test: record the credential kubectl sent when an EKS call comes back Unauthorized The IRSA pod subtest's first kubectl is the gateway-CA configmap, which lands before the pod dumps are registered, so a 401 there left no evidence at all. A verbose read-only probe reports whether an Authorization header was sent, which separates a token the apiserver rejected from one the exec plugin never made. * fix: report a starting engine promptly, and budget CI jobs from the suite The agent waited a full heartbeat interval before its first beat, so a restart could not be observed as finished for up to 30 seconds after the engine was already serving. Beat at startup instead, and report every 5 seconds until the engine has answered once — the control plane is actively waiting on that beat, and the steady-state cadence it hands back is for a database already serving. The rds suite is the one most sensitive to what else the hypervisor is doing: its DB VMs serialise behind a 4 GiB semaphore, so its wall clock is set by how long it waits for the budget. The same pass ran 45m beside four nightly cells and overran 50m beside nine, panicking with two tests still running. Raise its budget to 60m. That budget was then larger than the jobs meant to contain it. A job cancelled on timeout reports only "The operation was canceled" and takes the artifacts with it, so the inner `go test` timeout must always fire first — and it could not for rds on e2e.yml, nor for storagefault or diskperf, which have asked for 90m and 180m inside a 60m job since they were added. Both workflows now resolve their job budget from suite-sets.sh rather than carrying a literal: e2e.yml from the largest requested suite, e2e-suites.yml per matrix leg, so a leg is bounded by its own suite and not by the longest one beside it. The default push set resolves to exactly the timeouts both workflows had before. * fix: check a permutation cell's job budget against the suites it runs The guard only saw cells that name a suite, so the nine that run a topology's nightly subset were never checked against it — the same class of mistake the guard exists to catch, in the majority of the cells. Bound each by the longest suite in its subset: max() and not sum(), because whichever suite hangs must panic on its own timeout first, and a sum would demand a budget the set has never come close to using. Every cell clears it today. * fix: report why a PostgreSQL engine would not start, and stop failing a stop that landed Three defects, all found by the first rds pass that ever ran deep enough to reach them. A wait that saw an instance reach `failed` called t.Fatalf from inside the polling condition, and testify runs that condition on its own goroutine. The Goexit killed the tick and left the collector empty, which reads as the condition being satisfied — so the wait returned successfully on an instance it had just declared broken, logged "reached status available" about one that was failed, and let the test carry on against it. Six conditions across four suites call Fatal or require the same way. EventuallyErr now runs its condition on the test's goroutine, which makes all six behave as written. A DB VM stop has two drivers: the API call and the reconciler resuming one whose caller it assumes has died. When they overlap the loser is told the VM is not in a state to be stopped — which is the state the stop was heading for — and the customer's StopDBInstance returned InvalidDBInstanceState for a stop that had just completed. Both paths now treat a refused command the way they already treat an unanswered one: it says nothing about the VM, so the fleet view settles it. PostgreSQL had no equivalent of the engine log tail MariaDB has carried since it was added, so a postmaster that refuses a setting produced "engine did not respond on /run/postgresql:5432" and nothing else — on a guest with no SSH, whose console records only that pg_ctl could not start the server. The probe now quotes the tail of the log the image already directs the postmaster to, and the reboot and modify paths carry the agent's reason into the failure the way the create path always has. The reason reaches the customer through StatusInfos. The rds suite's budget goes to 90m and cell-27's job to 100m. Its wall clock is what the DB-VM queue costs rather than what the tests do — single tests logged 11m, 22m and 26m waiting for budget — and the pass above still had three of seventeen tests running at 60m. * fix: keep a DB VM recoverable when its agent cannot start, and flush the volume before a snapshot An rds-agent that failed to construct exited, so a transient IMDS credential resolution took the guest down for good: rds-datadir and rds-init cascade on the missing handoff, the engine never starts, and the control plane sees only silence until it declares the instance failed ten minutes later. Construction is retried like every other boot step instead. OpenRC skips start_post when the start itself fails, which is the one case with no diagnostic at all — the console carried nothing but "died". The log mirror moves out of the handoff wait so a failed start uses it too. A snapshot quiesced the engine and nothing else. The engine repairs its own files from the WAL, but postgresql.conf sits beside them and it does not: a restore came up on a torn tail and the postmaster stopped on a syntax error at the unterminated include hook. The quiesce now flushes the whole data mount, and the hook is renamed into place rather than appended, so a boot that finds a half written one replaces it instead of appending after its open quote. * fix: flush the config cloud-init wrote before a reset can take it A DB instance rebooted seconds after it went available came back with no RDS_GATEWAY_URL, so rds-agent could not reach the control plane and the engine never started. cloud-init writes agent.env once, per-instance, and never again; a reboot is a QMP system_reset with no guest shutdown, so the file was still dirty page cache when the machine reset under it and the next boot read nothing. The handoff wait now syncs, which is the point at which the agent has proved it read that file. * fix: stop dropping an EIP teardown published while vpcd is restarting DisassociateAddress publishes vpc.delete-nat fire-and-forget. On run 34046316024 vpcd disconnected from NATS at 03:51:26.73 and its replacement subscribed at 03:51:27.09; the publish landed at 03:51:26.92, in the gap, and went nowhere. ReleaseAddress then returned the address to the pool two minutes later with its /32 host route still delivering to the guest that had held it. vpc.delete-nat now uses request-reply like vpc.add-nat, and retries only on no-responders — which returns immediately, so the cost is paid solely when nothing is listening. A teardown that still fails is logged as an error naming the address rather than lost silently. Neither is fatal to the caller: the reconciler's orphan sweep remains the backstop. vpc.add-nat on the EIP path stays fire-and-forget. The shared helper blocks it on the flows barrier for up to 45s, and associating an address should not inherit that. Two reconcile passes in that run had the address absent from intent and did not sweep its host route, and the journal cannot say why: the sweep logged only when it acted, so a route it kept was indistinguishable from one it never saw. It now logs the bindings it found, kept and wanted. * ci: render one run-level table for the e2e nightly Reading a nightly meant opening nineteen jobs, and the Actions UI does not separate a cell that failed from one cancelled out from under it — a distinction every investigation of this suite has needed, and one the run 34046316024 analysis had to reach for the jobs API to get. A final job, needs: every cell and if: always(), renders cell, topology, result, duration, runner and per-cell test counts into the step summary, with run-level totals, the three slowest cells and the failing test names. Result, duration and runner come from the jobs API so a cell that died before writing junit still gets a row. Counts come from a new junit-only artifact per cell rather than the full cell artifact, so the roll-up does not pull nineteen journals it has no use for. Verified against run 34046316024: 14 success, 2 cancelled, 1 failure, with the two cancellations shown as cancelled rather than folded into failure. * fix: assert EC2's actual EIP semantics, and probe EKS auth on an authenticated path natuplink asserted that an instance's auto-assigned public IP is still plumbed after an EIP associated to the same instance is torn down. EC2 does the opposite: associating an EIP releases the auto-assigned address back into the pool, and it cannot be reused. Our datapath already does that — run 34059879666 shows 192.168.0.72's ingress removed at 07:09:44.000, 36ms before .73's was installed — so the test was asserting behaviour neither AWS nor we have. It now asserts the release at the point it happens, and that no public delivery remains after the teardown. This surfaced only once the teardown itself started working: the run's other assertions passed, including .73's removal, which was the failure before. Our records still disagree with our datapath — the ENI keeps the released address, which is why termination tears the old one down twice and never the current one. That is the remaining half of mulga-7zc9w and is not fixed here; it needs an IPAM release on the associate path. Separately, AuthDiagnostics read /version, which is anonymously readable, so its 200 was consistent with both branches it claimed to distinguish. It now runs auth whoami, reads kube-system, and invokes the credential plugin directly, reporting exit code, stderr and whether a token was produced — never the token, since CI logs outlive it. * test: capture per-guest state when a spread guest never answers SSH TestMultinodeSpread's only failure dump is dumpPlacementNATDiag, which collects OVN, OVS, ARP and route state aimed at the NAT and EIP half of the test. The failure it actually hits is a guest that never accepted SSH — run 34040324968, the guest on node3 at 10.100.2.5 — and for that it recorded no console, no port binding and no tap state. There was no way to tell from the artifact whether the guest booted at all. Now dumps, per guest: the console, the NB logical switch port, the SB port binding, the OVS interface backing its tap, and br-int's port list. Probed on every node rather than the expected one, because a port bound on the wrong chassis is exactly what a hosting-node-only dump cannot show. Covers all three guests, not just the failing one, so a working guest is available as the control. * ci: sample hypervisor load either side of every cell Eight runner names map onto three hypervisors, so cells that never share a runner still share a host: when cell 10's SSH wait expired on run 34040324968, cells 22, 17, 18 and 28 were concurrently live, two of them multi-node. Nothing in the artifacts measured host load, which left co-tenancy as an explanation that could be neither confirmed nor ruled out. Records loadavg, MemAvailable, SwapFree, the cpu line carrying steal, and the QEMU process count before and after each cell, into the cell artifact. * ci: note that main's missing concurrency block is why the nightly runs ungrouped Three refs carry three different groupings: main has no concurrency block at all, dev uses e2e-nightly, this branch uses e2e-hypervisors. Scheduled runs load the workflow from the default branch, so every cron nightly runs ungrouped and cannot queue against anything — which is how a dispatch of 34050630627 ran alongside the scheduled 34046316024 and cost it cells 24 and 27. e2e-hypervisors is the right group and deliberately broader than dev's, since e2e.yml, e2e-suites.yml and e2e-ddil.yml draw on the same runner pools. Nothing on a feature branch can fix this; the block has to reach main.	2026-09-06 16:53:32 -05:00
Ben Duncan	828b5df4aa	fix: pause a guest on a backend I/O error instead of handing it EIO (#977) ... * fix: fail DRAIN when a volume seal failed shutdownAndUnmount logged the volume unmount error and returned nothing, so StopAll returned nil and the DRAIN phase ACKed success with vms_stopped set. The coordinator then advanced to STORAGE and stopped viperblock while the seal was still in flight, tearing the VBState envelope and leaving the guest unable to mount its root volume on the next boot. The deploy reported success. shutdownAndUnmount now returns ErrVolumeSealFailed when the unmount did not seal; QMP, PID-file, fw_cfg and telemetry cleanup stay best-effort. stopCleanup and stopOne carry it without aborting the rest of the teardown, since QEMU is already down by then, and StopAll aggregates the per-VM failures so handleShutdownDrain ACKs with Error set and never reports vms_stopped. volumeMounterAdapter.Unmount was the second swallow site: it logged each failed seal and returned nil unconditionally. It now attempts every volume and returns the aggregate, which is what the VolumeMounter contract already described. Terminate still tolerates a failed seal — the volumes are deleted next, so an unsealed block map loses nothing a caller can ask for. This does not fix the seal itself. It converts silent data loss into a refused upgrade that an operator can act on. * test: add the storagefault e2e suite for backend-outage guest survival Freezes predastore with SIGSTOP to hold connections open while answering nothing, which is the fault a timeout-free path cannot distinguish from a slow peer. On one node that is a total outage; on three it freezes every node but the guest's own, breaking the RS(2,1) shard floor. The suite writes and flushes a verifiable region before the fault and churns a separate unverified region during it, so an interrupted write is never mistaken for lost data. The root-filesystem load runs from the user's home rather than /tmp, which is tmpfs on this image and would move no bytes on the root device. * test: verify the root device is under write pressure, and size the outage The previous guard matched /^ *vda / against /proc/diskstats, where the device name is the third field, and read $7 which is ms_reading rather than sectors_written. It printed empty and asserted nothing, so a green run could not distinguish a surviving guest from an idle disk. It now samples sectors_written twice and fails unless it advances. The outage is sized from viperblock's absorption rather than guessed: a workload must fill DefaultMaxPendingBytes before backpressure engages at all, so a short freeze measures the retry budget rather than the fault. * test: add nbdkit fault injection to the storage-fault e2e suite The predastore freeze takes the whole storage cluster away from every guest on it, which is too broad to run beside anything else. Removing a single volume's NBD server is the same fault at one-volume scope: one host, one guest, nothing cluster-wide stopped. Both signals are here because they sit on opposite sides of the defect. SIGSTOP holds the socket open and answers nothing, so the guest blocks and no error reaches ext4. SIGKILL closes it, and QEMU's nbd driver at the default reconnect-delay of 0 fails requests at once, which under werror=report aborts the journal. Signalling is guarded: a pid is matched on comm, our plugin and the volume argument together, and re-confirmed immediately before the signal, because these hosts run nbdkit for other volumes and pids are reused. * fix: pause a guest on a backend I/O error instead of handing it EIO Every virtio-blk device was created with werror=report,rerror=report, so a backend outage was delivered to the guest as EIO. ext4 aborts its journal on that and remounts read-only, and the write that failed is simply gone. Under werror=stop QEMU pauses the VM and holds the failed request, so a backend that comes back resumes the guest with nothing lost. The boot drive, the cold-boot data volume and the hot-plugged data volume all move to stop. Alongside it the NBD transport gets reconnect-delay=30, which pauses requests rather than failing them while a storage process restarts on the same socket; on the legacy -drive form it has to be file.reconnect-delay, because the bare name reaches the raw format driver, which has no such option. The QMP heartbeat now recognises run state io-error and issues cont. After ten attempts the instance is reported impaired and the cadence drops to one attempt per ten heartbeats, but it never stops: QEMU holds the request for as long as the guest exists, so a backend that returns an hour later can still be resumed. The storagefault suite gains the fault this was written for — kill the volume's own nbdkit, hold, then put it back on the same socket — and it now verifies a sha256 region across the outage rather than trusting the guest to be reachable. Three harness defects went with it: the socket-liveness poll ran unprivileged against a 0770 directory and could never succeed, the relaunched socket was left 0755 where QEMU needs group write, and the two variants shared one guest so a fault in the first was diagnosed in the second. * test: cover the io-error resume policy * test: stop two parallel tests racing on the terminate backoff variable * fix: report a guest paused on a backend I/O error as impaired A guest held by werror=stop kept reporting ok on both status checks. QMP stays responsive while the guest is paused -- cont is being sent and answered -- so the QMP failure gate that DescribeInstanceStatus keys on never trips, and the one state a customer most needs to see looked healthy. Instance status now reports impaired/failed while the resume budget is non-zero, timestamped from IOErrorSince so impairedSince is when the guest paused rather than when it was escalated. System status reports impaired too: a backend I/O error is an infrastructure fault, not a guest one, and that distinction is what the two checks exist to draw. The counter is cleared by the first poll that sees the guest running, so both clear on recovery. * test: reap the relaunched nbdkit when the fault test ends relaunchNbdkit detaches the server with setsid so it outlives the SSH session that started it, and viperblockd never learned it exists. Nothing reaped it, so every run left an nbdkit holding a socket and a cache for a volume that was later detached or deleted -- four were found running on dev-prod against two detached and two deleted volumes. The PID is resolved from the socket path rather than the volume id, so a legitimate viperblockd server for the same volume is never the one killed. * test: prove the backend is serving before a fault is called cleared Thawing the process a fault stopped is not the same as restoring the service. The outage test treated SIGCONT as the end of the fault and went straight to judging the guest, so it passed while the cluster it ran on stayed unable to serve a single object for the next 35 minutes, and the next test would have inherited that. thawPredastore now puts and gets an object through every gate, with a budget, and fails the test if the backend does not come back. * fix: do not hand back a mount entry whose export is gone A volume whose unmount failed kept its entry in MountedVolumes while its nbdkit was gone. Every later ebs.mount then matched that entry and returned its socket path, so vmMgr waited for a socket nothing would ever create and the instance could not be started again — on dev-prod a guest stopped against a frozen backend was left permanently stopped, four start attempts failing identically on a socket the node had already removed. Drop the entry when there is positive evidence the export is dead: the socket is absent, or the recorded pid is not running. Neither can be true of a live export, so the double-writer hazard the guard exists to prevent is unaffected. Also make the durability test's pattern helpers report failures. The read dd sat in a pipeline, so its exit status was discarded and the harness hashed nothing and returned the empty-string sha256 as a checksum. * test: report the guest console when a restart never answers SSH A restart that reaches running but never comes back is indistinguishable from a dead root filesystem without the serial console, and the log is gone once the fixture terminates the instance. * fix: converge guest port datapaths concurrently, not one deadline after another Each guest port that misses its SB binding burns the full convergence deadline, and applyEIPs spent them in series. Three ports belonging to long-stopped instances therefore held a newly launched guest's EIP dark for the sum of them before its own probe started. AddEIP stays serial since it mutates NB. Only the waits fan out, bounded, and sb-cluster-state-reset is now serialised with a minimum interval so concurrent probes cannot each fire a cluster-wide resync. * fix: refuse a stale cross-node start after the owning node is gone The volume lease already refused a cross-node mount, but it is a liveness claim: 45s TTL, renewed every 15s. A node that dies holding un-uploaded writes stops renewing, and 45 seconds later nothing refuses a start elsewhere that loads a stale backend checkpoint. A failed seal now writes a durable marker naming the node holding the only current copy, cleared on a seal that reaches the backend and on delete. The lease covers a live holder, the marker a dead one. The refusal also reaches the caller instead of being flattened to ServerInternal, which told an operator to retry something that cannot succeed while the holder is down. * fix: keep the daemon's message when a start is refused Rebuilding the error from the code alone rendered the generic ErrorLookup text, so a refusal naming the node holding the volume reached the caller as boilerplate about attaching volumes. * fix(ec2): terminate detaches every DeleteOnTermination=false volume TerminateStoppedInstance skipped the detach for a non-Boot DoT=false volume on the assumption that Stop's Unmount had already cleared its attachment. Unmount only clears it when the seal succeeded, so a volume whose seal failed stayed in-use pointing at an instance that terminate then deleted, leaving it unreachable with no owner to release it. The running-instance path already detaches every DoT=false volume; the stopped path now matches it, and tolerates an already-gone volume the same way. DeleteVolume also keeps its provider error on the way out. The terminate reaper retries the delete every two minutes and logged a bare ServerInternal, which made a blob node being down indistinguishable from a transient fault. * fix(ec2): DetachVolume --force releases a volume whose instance is gone The ordinary detach routes to ec2.cmd.{instanceID} for the guest unplug, so a volume left attached by a part-way teardown could never be detached: the instance no longer runs anywhere and nothing answers. That also made it undeletable, since DeleteVolume refuses an attached volume. ForceDetachVolume already existed for this deadlock but was reachable only from account teardown. Force now falls back to it when the instance has no responder, which is the one case where clearing the attachment cannot strand a live guest. * fix(viperblock): unsealed writes steer placement instead of pinning a volume The dirty marker refused any mount from a node other than the one holding unsealed writes. Instance start already forwards to the node that last ran the instance and falls back after a 30s window, so that refusal turned a working fallback into a volume that could run nowhere once its owner was gone. It traded a rare data-loss risk for a permanent availability failure. The marker is now a record and a warning, not a gate. A takeover logs the previous holder and why it could not seal, then adopts the marker so the old node is no longer treated as the better source. Live exclusion is unchanged and still enforced by the lease. It is also written on open rather than on seal failure. A node killed mid-write never reaches its seal, so the old trigger left no trace in exactly the case the marker exists to describe. Adds 'spx admin volumes unsealed' to see which volumes would cost something if started elsewhere. * fix(viperblock): fence a volume whose lease moved instead of logging it The lease was checked once, at mount, and never again. A node that lost it kept its export up and kept writing into a volume another node owned, so both copies advanced and the backend settled it by arrival order. Losing a renewal now decides: another holder means tear the export down, no holder means the entry merely aged out, so reclaim it and leave the guest running. Fencing never seals — this node's copy is the stale one — and leaves the dirty marker for the winner to take over. This is not an epoch fence. It narrows the window from the 45s TTL to one renewal interval at no write-path cost; the backend still cannot refuse a stale writer. * fix(network): advertise EIPs so a recycled address is reachable The gateway switch port only asked ovn-northd to publish the router's NAT addresses in centralised and routed modes, so on a distributed cluster nothing ever announced an elastic IP. OVN still answers ARP for it, which is why a first-time allocation works and hid this — but no gratuitous ARP is sent when the address moves to a new ENI, so every host on the segment keeps resolving it to the previous holder's MAC until its cache ages out. Reusing an address from the pool then lands on a dead MAC and the instance is unreachable from outside. Distributed mode needs the option as much as the others: northd scopes each entry with is_chassis_resident(<vif port>), so the announcement leaves from the chassis running the instance rather than the gateway, which is what distributed NAT wants. Converge on the already-attached path too. AttachIGW returns early once the gateway port exists and nothing else rewrites it, so without that an existing VPC would stay silent for its whole life. Confirmed on dev-prod: setting the option moved three EIPs from unadvertised to carrying their correct external_mac, each scoped to its own ENI's port. * fix(viperblock): make the volume fence and its dirty marker ownership-safe Four corrections to the lease-loss fence, all found in review. Drop the reacquire path. Losing a lease and finding no successor was treated as a lapsed entry worth reclaiming, but leases.acquire returns the cached entry — already flagged lost, with a dead renewal goroutine — so the "reclaimed" lease renewed nothing and the node went on writing a volume it could not prove it owned. The two errors are not symmetric: a false fence costs one guest a restart, a false reclaim leaves two nodes writing one volume. Fence in both cases, and report which happened through the outcome attribute. Surrender a lease this node cannot confirm. Renewal only noticed a loss when the server rejected an update, so a node partitioned from NATS kept its export up while the TTL expired and another node took the volume. Track the last confirmed renewal and give the volume up once that ages past the point where the server may admit another owner. Bound the renewal call too, so a hung JetStream request cannot hold the goroutine past that point. Do not act on a kill that did not take. Both the fence and ebs.unmount removed the mount entry, released the lease and announced the result before knowing nbdkit had exited — the fence then reported a writer stopped that was still running, and the unmount sealed into a directory a live process was writing. Wait for the exit, and on failure leave the volume mounted and say so. An already-exited process now counts as success, which is what the caller asked. Give the dirty marker a generation. It recorded an owner and nothing to order two nodes' claims, so a node returning after a takeover could clear or overwrite the winner's marker and point placement back at the stale copy. Stamp the lease generation, refuse a write from an older one, and only clear a marker this node still owns. A marker that cannot be written now fails the open rather than leaving writes nothing records. * test(viperblock): cover the stall surrender and the unmount kill guard The two ownership cases the fence work left untested. A holder partitioned from NATS while the winner keeps it never gets an update rejected, so revision checking cannot see it — it has to give the volume up on its own clock. Pins that validity sits below the server TTL with room for a check in between, that a lapsed holder reports the loss as a stall rather than as a takeover, and that releasing a surrendered lease leaves the entry for whoever holds it now. The unmount side of the kill guard: sealing under a live nbdkit writes a state torn between two writers, so a kill that did not take has to fail the unmount and leave the entry mounted for a retry. * fix(e2e): keep the guest private key for as long as the key pair EnsureKeyPair computed the PEM path from the caller's directory but wrote the file inside the ensureOnce closure, which runs only on the first call. Callers passed a per-test artifact dir, and ArtifactDir removes that dir when the test passes — so the first test's success deleted the only copy of the private key while the key pair stayed registered, and every later call returned a path nothing had ever been written to. Guests after the first were then launched against a public key whose private half no longer existed, so the harness could never authenticate and gave up after the full five-minute budget. The guests were healthy and reachable on tcp/22 throughout, which is why this read as a storage or network fault for so long. The directory parameter was the defect, so it is gone: the PEM now lives in a directory the fixture owns and removes with the key pair, and no caller can supply one that is shorter-lived. Also reorder the killed-nbdkit teardown. The reaper registered its own cleanup after the detach registered one, so LIFO killed the block server before the platform detached the disk and DetachVolume then failed with "Node ... is in use". The detach cleanup now reaps once it is done. * fix: stop leaking an nbdkit export when recovery cannot adopt it Restart recovery adopts a guest's boot volume, then routinely fails the -efi volume with "claim lease: context deadline exceeded" -- JetStream is still catching up at daemon start and the KV claim runs on the client's 5s default. Recovery gave up on the first error and never retried, so the export stayed live with no registry entry. At terminate the unmount found nothing to seal, returned NotFound and killed nothing, and the process outlived its guest holding the volume and socket open. EFI is always the second volume adopted, which is why it was always the one that leaked. Two changes, because either alone leaves a hole: - rebuildMountedVolume retries the build (4 attempts, 3s apart) so a JetStream that is merely slow no longer orphans an export for the life of the process. - unmountVolume reaps a corroborated export for the volume it was asked to unmount when the registry has no entry for it, so an export orphaned by any means is still cleaned up. The reap goes through reapOrphanedNbdkit, which declines anything a live process references: at unmount the guest's blockdev may not be torn down yet, and killing the server underneath it strands QEMU. * test: hold the backend-serving probe and assert predastore readiness The backend-outage test declared the backend restored on one successful round trip through each gate. A gate can answer the request that arrives just after a thaw and then lose its blob peers under the load that follows, which is a backend that never recovered reported as one that did -- and that is what happened: the probe reported all three gates serving minutes before the guest was still taking BLOCK_IO_ERROR on every resume. The round trip has to hold across three rounds now, and the count restarts rather than decays when one fails, so a gate that serves and then stops is the failure the hold is looking for rather than noise it averages out. It also checks /readyz on each host and reports any failed check by name. The round trip only proves this workstation can write through a gate; it says nothing about which blob peers that gate can reach, and a gate one node short of the write floor serves a small object and fails a guest. A host that serves no admin listener is logged and skipped, since the port is optional configuration. * fix: scope predastore's admin port to the cluster plane in the host policy predastore's admin listener binds the host's cluster bind_addr and serves /healthz and /readyz unauthenticated. Readiness now names which blob peers a gate cannot reach, which is exactly the kind of topology detail 4432 is peer-scoped for, so 8660 joins it in the peer-scoped rule rather than the public one. * fix: keep resuming a paused guest at heartbeat rate after a long outage Past ten attempts the cont dropped to one per ten heartbeats, so a guest whose backend had come back could sit paused for another five minutes before anything retried it. Measured on dev-prod: the backend was proved serving through all three gates while the guest was still waiting on attempt 60, five minutes after attempt 50. The slowdown was there to avoid paying a failed I/O every thirty seconds on a backend still down. That is nearly free, and five minutes of downtime after recovery is not, so only the log line is throttled now. * test: fail the run when a key pair name resolves to more than one pair A create that lands twice leaves two key pairs with one name and different material. RunInstances resolves the name to one of them while the harness holds the private key of the other, and the only symptom is every guest in the run refusing the key at the SSH gate — five minutes of waiting per guest, four guests, and nothing anywhere naming the cause. Observed on dev-prod: two pairs created a millisecond apart, both named e2e-key-30de10ba. One DescribeKeyPairs after the create says so outright instead. * test: tolerate an already-released volume in the detach teardown Teardown runs DetachVolumeWait after a terminate that releases the volume itself, so the detach can arrive at a volume already on its way to available. That is the outcome it wanted, and treating the IncorrectState refusal as fatal failed a backend-outage run whose assertions had all passed — the guest came back, every acknowledged write verified, and the test still went red on the way out. VolumeInUse is still fatal: that one means the volume is attached somewhere it should not be. * test: carry the merged suites onto the reworked key-pair and account-scoped seams The diskperf suite arrived on dev calling EnsureKeyPair with the artifact dir this branch removed, and DeleteVolume's metadata lookup is account-scoped now, so the provider-failure test has to seed and delete under the same account. * refactor: make the fence outcome a type in one const block Four separate exported string constants with a paragraph of comment each, passed to RecordVolumeFence as a bare string. Group them, give them a FenceOutcome type, and take that type at both call sites so an arbitrary value can no longer reach the metric attribute. * refactor: move viperblockd's wire contract into a leaf vbwire package The daemon imported viperblockd for exactly two symbols -- the fenced volume subject and its event -- which put the storage service back on the daemon's dependency graph to name something it only listens for. vbwire holds the contract instead: the subject and event, the dirty bucket name and record shape, and the read-only listing an operator tool needs. It imports nothing from spinifex, so both the daemon and the admin CLI drop their edge to viperblockd entirely. cmd/service.go keeps its one, which is genuine: it constructs the service. * docs: drop a specific host name from the start-refusal comment Named a real dev-prod host in a comment and in test fixtures. Say what the name is instead. * ci: give storagefault a dedicated nightly cell The suite is eligible for both topologies and runs on the PR gate, but no nightly cell ran it: the permutation cells run a deliberately narrow subset and none of the dedicated suite cells named it. Cell 29, single-node, following eks/ecs/rds. It cannot join a permutation cell -- it SIGSTOPs a cluster-wide service, so every suite sharing that environment would report the outage as its own failure. * ci: run storagefault nightly on three nodes, not one The previous commit put its dedicated cell on a single-node runner. That was the wrong topology and the wrong justification: single-node is RS(1,0), where the shard floor cannot be broken and a cross-node takeover after a failed seal does not exist, so the two assertions worth having a dedicated cell for were the two it could not make. Cell 29 is now real-multi on ci-multi. It also corrects the claim that the PR gate already covers the multi-node case. The suite is eligible for both topologies, but e2e.yml's default suite list never names it, so it has never run there. A full pass is ~63 minutes on dev-prod bare metal, almost all of it deliberate waiting -- one test holds predastore SIGSTOPped for seven minutes -- which is too long to add to a gate that currently finishes well inside its budget. Nightly is where it belongs. Consequences: a per-cell timeout_minutes, defaulting to the previous 50, since 90 minutes of suite cannot run in a 50-minute job; the permutation steps gated on matrix.suite == '' so the dedicated cells do not also run them; and a Run Multi Suite E2E step that builds NODES_CSV from the tofu inventory and asserts the suite against E2E_SUITES_MULTI before running it. * fix: stop a startup KV race leaving vpcd serving keyless guests vpcd opens five KV buckets while NATS is still coming up, each attempt carrying a ~5s JetStream API deadline, and applied three different policies to the same failure: spinifex-vpc-enis was fatal, while instance-state, security-groups, subnets and vpcs each degraded silently for the lifetime of the process. A five-second blip during a cold start therefore decided how a node served metadata until someone restarted it. On dev-prod all three nodes hit this after a deploy. Two were saved by accident: their ENI bucket also timed out, so they exited 1 and systemd restarted them cleanly. The third's ENI bucket opened in time, so it never crashed and ran for hours with no instance record loader. With the local state file unreadable as well, every instance-backed IMDS leaf 404'd, cloud-init's crawl died on ami-id and fell back to DataSourceNone, and every guest launched on that node booted with no authorized_keys. kvstore.OpenWithRetry retries for a bounded window before any policy applies, logs the cause per attempt and the wait on a late success, wraps the last error with the bucket name, and returns at once on a cancelled context. All five opens use it. Failing to open the instance record space is now fatal, matching the ENI bucket: the dependency whose loss stops guests logging in should not be treated more leniently than the one whose loss stops ENI resolution. Also fixes the reason the fallback was load bearing at all. The local state file was written 0600 and vpcd reads it as a different user in the shared spinifex group, so the local-first path had never once worked and its mtime+size cache had never fired. It is 0640 now, set with an explicit Chmod so an inherited umask cannot undo it. Closes mulga-gtpzz, mulga-dhqxr. Still open: mulga-mm2oc, IMDS serving an unresolvable instance as "no key pair" with nothing logged. * fix: scope the ime-* firewall accept to IMDS and VPC DNS only The per-tap OVS demux delivers any guest packet addressed to 169.254.169.254 or .253 onto an ime-* host port regardless of L4 port, and iifname "ime-*" accept then admitted all of it. Every wildcard-bound host service — sshd, the AWS gateway, northstar, the console — answered guest packets on those addresses; only asymmetric reply routing kept the handshake from completing, and a guest could still drive northstar to recurse unthrottled. Scope the accept to IMDS (.254:80) and VPC DNS (.253:53), letting everything else fall through to policy drop. * docs: document the instance-to-host plane and how to protect it without the ISO The metadata/VPC-DNS link-local plane (169.254.169.254/.253) is served by host services and is invisible to a NIC-level or cloud firewall, since the traffic is intra-host on the OVS bridge. Explain that the ISO ships the scoping rule armed, that the binary installer ships the firewall off, and that a non-ISO install must enable it with setup.sh --firewall=on or firewall_enabled=true. Include verification from an instance and the host. * fix: withhold the IMDS meta-data listing when the instance does not resolve Every omission from the meta-data root is a positive claim about the instance — public-keys/ absent means "launched without a key pair", and cloud-init believes it, writes no authorized_keys and boots the guest keyless. So resolving the instance was best-effort for a listing whose whole meaning depends on it: a backend error or a miss shortened the listing instead of failing it, and nothing anywhere recorded that the answer was a guess. Withhold the listing with a 500 instead. cloud-init retries a 5xx and cannot retry a 200 that is merely wrong, so this converts a permanent silent failure into a transient loud one. servePublicKeys already had this policy for the leaf; the root is what contradicted it. A profile lookup that fails still serves the listing minus iam/, and now logs. The key pair is the field a guest cannot recover from losing, while a missing role surfaces through the credentials leaf, which reports the miss itself. Also stops the storagefault cross-node test failing on its own access path. It tolerates the pattern changing across a takeover, but readPattern needs sudo, and a volume opened from the backend checkpoint may predate cloud-init writing its sudoers drop-in — indistinguishable from an unreadable device at the dd call site. It now checks sudo first, and reads cloud-init's recorded datasource to tell a rolled-back root filesystem from IMDS never having answered, failing loudly on the latter rather than accepting it as storage loss. --------- Co-authored-by: brynmailer-mdc <bdm@mulgadc.com>	2026-09-03 19:45:33 -05:00
Bryn Mailer	18aa67b0e9	feat: add the e2e guest disk-performance gate (#964) ... * feat: add the e2e guest disk-performance gate Nothing measured guest disk performance in CI, which is why a quadratic index scan shipped in viperblock 35dd9f2 and survived three subsequent fixes to adjacent code. It is also why the two availability clauses at the end of the ebs-performance plan's Stage 0 have never been evaluated: a sustained random-write workload took the guest, and then the host, down, and nothing since has demonstrated that it no longer does. tests/e2e/diskperf/ launches an instance, attaches a dedicated blank volume per repetition, and drives a pinned fio profile over guest SSH. Two profiles: randwrite-4k at 4 jobs x 4 GiB, the shape the exit criterion names, and randrw-70-30-4k, the mixed workload the headline IOPS came from. The target is a raw device with direct=1 rather than a file on the boot disk, so guest page cache, filesystem allocation and journal traffic are out of the path. Two classes of assertion, kept apart because folding them together lets a run that left the host unreachable pass on a respectable IOPS figure. Availability is hard pass/fail and absolute, so it gates from the first run: the job completes inside its budget, a guest responsiveness probe answers throughout on a second SSH channel, and the host's own sshd answers throughout. The guest probe reads the root volume while the workload runs on a different one, so a pass is evidence of isolation between volumes rather than of the job merely being slow. The host probe is a TCP connect plus an SSH banner read, which needs no credential and still proves the kernel scheduled sshd. Throughput and latency are banded against a committed baseline: fail above 25% regression, warn above 10%, per-metric median across repetitions. baseline.json is committed and embedded, with every figure zeroed, and a zero reference records and warns rather than failing. Only the availability class gates until numbers are captured on the bare-metal cell; a throughput band nobody measured would be a fabricated gate. A fresh volume per repetition is the reset. Extent count grows monotonically, so reusing a volume makes the gate order-dependent and drifts its own numbers upward over time, slowly normalising the regression it exists to catch. Each job is its own guest invocation and fio's JSON is fetched by a separate command, so a wedged job loses only itself and reports as a budget breach rather than a parse failure. e2e-baremetal.yml gains suites, suite-env and cell inputs so one reusable workflow serves both the correctness cell and nightly cell 28, which queues behind cell 20 on the shared concurrency group. Cells 25-27 are the dedicated eks, ecs and rds cells. The per-suite Go test timeout is resolved from e2e_suite_timeout in suite-sets.sh, where diskperf declares its 180-minute budget beside rds's rather than in a second place. The gate's own arithmetic runs in preflight via test-harness: band edges, the unmeasured cases, per-stream metric selection, fio JSON parsing, and the probe accounting including a slow-but-successful answer counting as a breach. A gate whose comparison logic is untested is the failure mode the plan is otherwise full of. The extent-count ceiling named as a fourth availability assertion is not implemented: it needs viperblock.index.entries from S1.5, which does not exist. * feat: let a manual bare-metal run choose its e2e suites A dispatch run only sees inputs declared under workflow_dispatch, which had none, so suites and suite-env resolved empty and fell back to the correctness set. Capturing a diskperf baseline by hand was therefore unreachable without a workflow_call caller. * fix: build the bare-metal suites the run step will execute The build step packs pre-built binaries into the tarball and did not see the suite selection, so a selected suite was absent from the node unless it happened to be one of the three the build script hardcoded. * feat: commit the measured disk-performance baseline Medians of three repetitions on the bare-metal cell. The throughput and latency bands now gate instead of warning, so a 2x regression in either job fails the suite. randwrite-4k's write tail is two orders above the mixed job's. It is recorded as measured rather than zeroed: an ungated metric would miss the regression class the gate exists to catch.	2026-09-03 14:33:21 +10:00
Julian Sommer	4522494c17	fix: define the per-suite e2e test timeout once, in suite-sets.sh ... The timeout was written out separately in each workflow and drifted: rds ran 50m on the PR gate and 30m nightly, where it panicked and took six in-flight subtests with it. e2e-suites.yml had the same hardcoded 30m on its single path, which also runs rds. A new e2e_suite_timeout helper sits beside the eligibility lists it already guards, and all four call sites resolve through it. e2e.yml loops over suites on the remote VM, which has no checkout, so it resolves the pairs on the runner and looks them up in the loop. Also lift the nightly cell ceiling to 50 minutes so the rds cell is not cut short by the job before its own timeout applies.	2026-09-03 09:21:17 +10:00
Julian Sommer	be94beb52a	ci: run eks, ecs and rds nightly in dedicated cells (#984)	2026-09-02 16:16:05 +10:00
Ben Duncan	f41f5530a0	test: add an e2e suite proving quotas gate every capped resource	2026-08-20 02:26:27 +10:00
Julian Sommer	5e17169661	feat(e2e): complete the RDS end-to-end suite (#608)	2026-08-03 15:38:07 +10:00
Josh Siviour	68abc0bfb3	fix(ci): scope e2e telemetry per cell/job and define suite topology eligibility once (#573) ... * feat(ci): extend per-cell telemetry scoping to every e2e workflow ci.cell only reached e2e-nightly and e2e-suites, so a run of e2e, e2e-ddil or gpu-e2e still landed in the sink as one undifferentiated bucket. Stamp it on the remaining workflows: single/multi on e2e's two provisioning jobs, ddil and gpu on their single-job workflows. The best-effort numeric job id lookup moves out of e2e-nightly into a shared composite action, so the four other provisioning jobs get ci.job_id without copying the gh/jq incantation. The lookup now matches on env.JOB_NAME rather than interpolating the name into the jq filter, which broke on names carrying quotes or non-ASCII separators, and no-ops when telemetry is disabled. gpu-e2e gets ci.cell but deliberately no job id lookup: nothing in that job bakes the attrs into a daemon, so the API call would be paid for an env var no exporter reads. * fix(ci): define e2e suite topology eligibility once Three workflows carried three different answers to which suites a topology can run. e2e-suites and e2e-nightly ran lb on their single-node leg, where every internet-facing subtest skips for want of a peer, so a fraction of the suite reported as a green lb run. e2e-suites' multi leg ran multinode alone, dropping the lb and cert coverage that only a multi-node env can complete — so full lb coverage never ran there at all. .github/scripts/suite-sets.sh is now the single definition, sourced by all three. e2e.yml's two allow-list regexes derive from it, e2e-suites derives both matrices from it (multi_provision gains the suite-matrix step single_provision already had), and e2e-nightly's deliberately narrower per-cell lists are declared there too, guarded by an assert that fails the cell if a subset ever names a suite its topology cannot fully exercise. The nightly stays narrow on purpose: its cells have a 35-minute budget and exist to prove each install/network/host-OS permutation boots and serves. Suite breadth is the PR gate's job and already runs on every push. That is now stated in one place rather than implied by a third divergent list.	2026-07-22 17:18:55 +10:00