-
Notifications
You must be signed in to change notification settings - Fork 3.8k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
storage: release quota on failed Raft proposals #38632
storage: release quota on failed Raft proposals #38632
Conversation
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
thanks for tracking this down, my only substantial comment is about where to defer cleanupFailedProposal
. Looking forward to a flurry of close notifications in my inbox.
A for of the early issues
Something's off there.
Reviewed 2 of 2 files at r1, 1 of 1 files at r2, 1 of 1 files at r3, 3 of 3 files at r4, 4 of 4 files at r5.
Reviewable status: complete! 1 of 0 LGTMs obtained (waiting on @nvanbenschoten)
pkg/storage/replica_proposal_buf.go, line 417 at r4 (raw file):
proposals,
pkg/storage/replica_raft.go, line 239 at r5 (raw file):
This ensures
pkg/storage/replica_raft.go, line 241 at r5 (raw file):
// proposal buffer successfully. This ensure that we always release any // quota that we acquire. defer func() {
Is this in the right place? I'd have thought this would pair more naturally somewhere around
cockroach/pkg/storage/replica_raft.go
Lines 190 to 192 in f39dd66
if err := r.maybeAcquireProposalQuota(ctx, proposal.quotaSize); err != nil { | |
return nil, nil, 0, roachpb.NewError(err) | |
} |
i.e. just after acquiring the quota. As written, we have to worry about any early returns between acquiring the quota and calling into propose
. For example, and this doesn't matter in production but demonstrates the point, a TestingProposalFilter
right now will leak any acquired quota:
cockroach/pkg/storage/replica_raft.go
Lines 201 to 206 in f39dd66
if pErr := filter(filterArgs); pErr != nil { | |
return nil, nil, 0, pErr | |
} | |
} | |
maxLeaseIndex, pErr := r.propose(ctx, proposal) |
Release note: None
Release note: None
This avoids confusing arithmetic in a few spots. Release note: None
This commit makes propBuf.FlushLockedWithRaftGroup more careful about accidentally dropping proposals when flushing its array. In most cases, any error will result in a log.Fatal call further up the stack anyway, but it's good to be careful about this stuff or we risk leaking proposals. We now make sure to register all proposals with the proposer after an error, even if though we stop attempting to propose entries with the Raft group. Release note: None
Fixes cockroachdb#34180. Fixes cockroachdb#35493. Fixes cockroachdb#36983. Fixes cockroachdb#37108. Fixes cockroachdb#37371. Fixes cockroachdb#37384. Fixes cockroachdb#37551. Fixes cockroachdb#37879. Fixes cockroachdb#38095. Fixes cockroachdb#38131. Fixes cockroachdb#38136. Fixes cockroachdb#38549. Fixes cockroachdb#38552. Fixes cockroachdb#38555. Fixes cockroachdb#38560. Fixes cockroachdb#38562. Fixes cockroachdb#38563. Fixes cockroachdb#38569. Fixes cockroachdb#38578. Fixes cockroachdb#38600. _A lot of the early issues fixed by this had previous failures, but nothing very recent or actionable. I think it's worth closing them now that they should be fixed in the short term._ This fixes a bug introduced in 1ff3556 where Raft proposal quota is no longer released when Replica.propose fails. This used to happen [here](cockroachdb@1ff3556#diff-4315c7ebf8b8bf7bda469e1e7be82690L316), but that code was accidentally lost in the rewrite. I tracked this down by running a series of `import/tpch/nodes=4` and `scrub/all-checks/tpcc/w=100` roachtests. About half the time, the import would stall after a few hours and the roachtest health reports would start logging lines like: `n1/s1 2.00 metrics requests.slow.latch`. I tracked the stalled latch acquisition to a stalled proposal quota acquisition by a conflicting command. The range debug page showed the following: <image> We see that the leaseholder of the Range has no pending commands but also no available proposal quota. This indicates a proposal quota leak, which led to me finding the lost release in this error case. The (now confirmed) theory for what went wrong in these roachtests is that they are performing imports, which generate a large number of AddSSTRequests. These requests are typically larger than the available proposal quota for a range, meaning that they request all of its available quota. The effect of this is that if even a single byte of quota is leaked, the entire range will seize up and stall when an AddSSTRequests is issued. Instrumentation revealed that a ChangeReplicas request with a quota size equal to the leaked amount was failing due to the error: ``` received invalid ChangeReplicasTrigger REMOVE_REPLICA((n3,s3):3): updated=[(n1,s1):1 (n4,s4):2 (n2,s2):4] next=5 to remove self (leaseholder) ``` Because of the missing error handling, this quota was not being released back into the pool, causing future requests to get stuck indefinitely waiting for leaked quota, stalling the entire import. Release note: None
23b1fb9
to
ba3813c
Compare
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Something's off there.
Done.
bors r=tbg
Reviewable status: complete! 0 of 0 LGTMs obtained (and 1 stale) (waiting on @tbg)
pkg/storage/replica_proposal_buf.go, line 417 at r4 (raw file):
Previously, tbg (Tobias Grieger) wrote…
proposals,
Done.
pkg/storage/replica_raft.go, line 239 at r5 (raw file):
Previously, tbg (Tobias Grieger) wrote…
This ensures
Done.
pkg/storage/replica_raft.go, line 241 at r5 (raw file):
Previously, tbg (Tobias Grieger) wrote…
Is this in the right place? I'd have thought this would pair more naturally somewhere around
cockroach/pkg/storage/replica_raft.go
Lines 190 to 192 in f39dd66
if err := r.maybeAcquireProposalQuota(ctx, proposal.quotaSize); err != nil { return nil, nil, 0, roachpb.NewError(err) } i.e. just after acquiring the quota. As written, we have to worry about any early returns between acquiring the quota and calling into
propose
. For example, and this doesn't matter in production but demonstrates the point, aTestingProposalFilter
right now will leak any acquired quota:cockroach/pkg/storage/replica_raft.go
Lines 201 to 206 in f39dd66
if pErr := filter(filterArgs); pErr != nil { return nil, nil, 0, pErr } } maxLeaseIndex, pErr := r.propose(ctx, proposal)
Done.
38632: storage: release quota on failed Raft proposals r=tbg a=nvanbenschoten Fixes #34180. Fixes #35493. Fixes #36983. Fixes #37108. Fixes #37371. Fixes #37384. Fixes #37551. Fixes #37879. Fixes #38095. Fixes #38131. Fixes #38136. Fixes #38549. Fixes #38552. Fixes #38555. Fixes #38560. Fixes #38562. Fixes #38563. Fixes #38569. Fixes #38578. Fixes #38600. _A lot of the early issues fixed by this had previous failures, but nothing very recent or actionable. I think it's worth closing them now that they should be fixed in the short term._ This fixes a bug introduced in 1ff3556 where Raft proposal quota is no longer released when `Replica.propose` fails. This used to happen [here](1ff3556#diff-4315c7ebf8b8bf7bda469e1e7be82690L316), but that code was accidentally lost in the rewrite. I tracked this down by running a series of `import/tpch/nodes=4` and `scrub/all-checks/tpcc/w=100` roachtests. About half the time, the import would stall after a few hours and the roachtest health reports would start logging lines like: `n1/s1 2.00 metrics requests.slow.latch`. I tracked the stalled latch acquisition to a stalled proposal quota acquisition by a conflicting command. The range debug page showed the following: ![Screenshot_2019-07-01 r56 Range Debug Cockroach Console](https://user-images.githubusercontent.com/5438456/60554197-8519c780-9d04-11e9-8cf5-6c46ffbcf820.png) We see that the Leaseholder of the Range has no pending commands but also no available proposal quota. This indicates a proposal quota leak, which led to me finding the lost release in this error case. The (now confirmed) theory for what went wrong in these roachtests is that they are performing imports, which generate a large number of AddSSTRequests. These requests are typically larger than the available proposal quota for a range, meaning that they request all of its available quota. The effect of this is that if even a single byte of quota is leaked, the entire range will seize up and stall when an AddSSTRequests is issued. Instrumentation revealed that a ChangeReplicas request with a quota size equal to the leaked amount was failing due to the error: ``` received invalid ChangeReplicasTrigger REMOVE_REPLICA((n3,s3):3): updated=[(n1,s1):1 (n4,s4):2 (n2,s2):4] next=5 to remove self (leaseholder) ``` Because of the missing error handling, this quota was not being released back into the pool, causing future requests to get stuck indefinitely waiting for leaked quota, stalling the entire import. Co-authored-by: Nathan VanBenschoten <nvanbenschoten@gmail.com>
Build succeeded |
Fixes #34180.
Fixes #35493.
Fixes #36983.
Fixes #37108.
Fixes #37371.
Fixes #37384.
Fixes #37551.
Fixes #37879.
Fixes #38095.
Fixes #38131.
Fixes #38136.
Fixes #38549.
Fixes #38552.
Fixes #38555.
Fixes #38560.
Fixes #38562.
Fixes #38563.
Fixes #38569.
Fixes #38578.
Fixes #38600.
A lot of the early issues fixed by this had previous failures, but nothing very recent or actionable. I think it's worth closing them now that they should be fixed in the short term.
This fixes a bug introduced in 1ff3556 where Raft proposal quota is no longer released when
Replica.propose
fails. This used to happen here, but that code was accidentally lost in the rewrite.I tracked this down by running a series of
import/tpch/nodes=4
andscrub/all-checks/tpcc/w=100
roachtests. About half the time, the import would stall after a few hours and the roachtest health reports would start logging lines like:n1/s1 2.00 metrics requests.slow.latch
. I tracked the stalled latch acquisition to a stalled proposal quota acquisition by a conflicting command. The range debug page showed the following:We see that the Leaseholder of the Range has no pending commands but also no available proposal quota. This indicates a proposal quota leak, which led to me finding the lost release in this error case.
The (now confirmed) theory for what went wrong in these roachtests is that they are performing imports, which generate a large number of AddSSTRequests. These requests are typically larger than the available proposal quota for a range, meaning that they request all of its available quota. The effect of this is that if even a single byte of quota is leaked, the entire range will seize up and stall when an AddSSTRequests is issued. Instrumentation revealed that a ChangeReplicas request with a quota size equal to the leaked amount was failing due to the error:
Because of the missing error handling, this quota was not being released back into the pool, causing future requests to get stuck indefinitely waiting for leaked quota, stalling the entire import.