Sharded WAL causes incomplete write errors in E2E tests. Disable by
default in hive.toml template, service.go fallback, and viperblock
constructor. Can be re-enabled via hive.toml shardwal=true.
Replace -serial pty with socket chardev + logfile for serial output:
- Console log written to {RuntimeDir}/console-{instanceID}.log
- Serial socket at {RuntimeDir}/serial-{instanceID}.sock for admin access
- Eliminates PTS extraction deadlock (mulga-lfo) — no more ptsChan
- GetConsoleOutput API handler reads log, returns base64-encoded output
- Per-instance NATS topic ec2.{instanceID}.GetConsoleOutput (no fan-out)
- Admin debugging via: socat -,raw,echo=0 UNIX-CONNECT:serial-*.sock
Update viperblock dependency to v1.0.13 which fixes CreateSnapshot
skipping WAL flush/consolidate when the VB instance doesn't own the
WAL files (viperblockd snapshot scenario).
Add local Docker E2E test instructions to TESTING.md.
Viperblock v1.0.12 fixes WriteShardedWALToChunk corrupting NBD plugin's
active WAL files when called from a VB instance that doesn't own the WAL
(viperblockd snapshot path). This caused CreateImage to fail with
"checksum mismatch in shard 0" on the second snapshot.
is_multinode() counted all [nodes.*] TOML subsections (daemon, nats, awsgw,
etc.) instead of just top-level node entries. A single-node config has ~8
such sections, causing the single-node Docker E2E to use coordinated cluster
shutdown and exit with code 254.
Fix grep to match only [nodes.X] (not [nodes.X.Y]) and preserve the original
test exit code in the cleanup trap.
Add [viperblock] section to hive.toml with shardwal config (default
true). The setting flows from config through viperblockd to nbdkit
plugin args, allowing operators to disable sharded WAL if needed.
Bumps viperblock dependency from v1.0.10 to v1.0.11 which includes
the sharded WAL implementation (UseShardedWAL/OpenShardedWAL), fixing
CI build failures.
With UseShardedWAL now defaulting to true, the EFI and cloud-init
volume preparation must open sharded WAL files instead of always
opening the legacy WAL. Without this, Flush() dispatches to the
sharded path but shard DB files are nil, causing zero-byte writes.
The FilterUnavailableTypesAfterAllocation test picked a random 2-vCPU
type from the raw instance type map, which could include memory-hungry
types (e.g. r5a.large at 16 GB) that exceed host RAM on CI runners
(~15.6 GB). Now picks from the already-filtered available types list.
Internal development plans, roadmap, and detailed dev process docs moved
to private mulga repo. CLAUDE.md trimmed to remove SSH credentials, node
IPs, and internal deployment workflows. Gosec fixes for file permissions
and unhandled error.
Julian Sommer