gsadmin/rustfs
1
0
Fork 0
mirror of https://github.com/rustfs/rustfs.git synced 2026-08-21 20:06:37 +00:00
Code Issues Packages Projects Releases Wiki Activity

feat(heal,scanner): best-effort repaired notices from the MRF consumer (#6283)

Browse Source 
The scanner's pending-heal ledger and the MRF journal tracked the same
damaged objects with no cross-talk: once the consumer landed an intent
with the heal manager, the ledger's retry entry for that target kept
re-submitting a heal the manager already owned (backlog#1894 axis B).

Fan the acceptance out: both dispatch sites in the MRF queue (the live
consumer and the startup replay) record a compact MrfRepairedEvent
(bucket, object, version bytes) in a bounded process-wide ring owned by
rustfs-common. The scanner drains its own bucket's notices at the top
of retry_pending_scanner_heals and clears the matching Object-kind
ledger entries in one batched retain + sync (a mass-recovery first
sweep must not turn into thousands of full-table ledger clones on the
scan task), with nil notice UUIDs mapping to None per the repo-wide
defensive-UUID invariant so unversioned entries match unversioned
notices only. Notices are best-effort by design — a lost or capped-out
notice leaves the entry to expire through its own attempts/age limits,
because the ledger is a retry oracle, not a source of truth; other
buckets' notices stay queued for their own scanners. Neither persistent
format changes; old nodes that keep double-booking remain harmless.

Co-authored-by: heihutu <heihutu@gmail.com>
This commit is contained in:
houseme
2026-08-20 01:31:26 +08:00
committed by GitHub
parent 99c3811d93
commit d6efb65588
4 changed files with 232 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
crates/heal/src/heal/mrf_queue.rs
+10 -3
View File
@@ -409,8 +409,13 @@ impl MrfRuntime {
let request = build_heal_request(&intent);
match manager.submit_heal_request(request).await {
// Accepted intents leave the pending set; the next flush persists the
// smaller snapshot, which is the journal's compaction.
Ok(HealAdmissionResult::Accepted) | Ok(HealAdmissionResult::Merged) => {}
// smaller snapshot, which is the journal's compaction. Fan out a
// best-effort repaired notice so retry ledgers (the scanner's
// pending-heal oracle) can drop entries whose repair the manager
// now owns (backlog#1894 axis B).
Ok(HealAdmissionResult::Accepted) | Ok(HealAdmissionResult::Merged) => {
rustfs_common::mrf_channel::note_mrf_repaired(&intent.bucket, &intent.object, intent.version_id);
}
Ok(HealAdmissionResult::Full) | Ok(HealAdmissionResult::Dropped(HealAdmissionDropReason::QueueFull)) => {
intent.attempts = intent.attempts.saturating_add(1);
if intent.attempts >= MRF_MAX_ATTEMPTS {
@@ -514,7 +519,9 @@ async fn replay_into(
while let Some(mut intent) = queue.pop_front() {
let request = build_heal_request(&intent);
match manager.submit_heal_request(request).await {
Ok(HealAdmissionResult::Accepted) | Ok(HealAdmissionResult::Merged) => {}
Ok(HealAdmissionResult::Accepted) | Ok(HealAdmissionResult::Merged) => {
rustfs_common::mrf_channel::note_mrf_repaired(&intent.bucket, &intent.object, intent.version_id);
}
Ok(HealAdmissionResult::Full) | Ok(HealAdmissionResult::Dropped(HealAdmissionDropReason::QueueFull)) => {
intent.attempts = intent.attempts.saturating_add(1);
if intent.attempts < MRF_MAX_ATTEMPTS {