kvserver: test Replica like raft.RawNode #105177

tbg · 2023-06-20T12:52:16Z

There are lots of "situations of interest" that we attempt to exercise by setting
up a TestCluster and an often complicated web of interceptors, all to
enact a desired "interesting" state of the system. This often involves a
delicate dance between wanting to get into this state quickly (not adding a
slow test) but also not having it flake due to excessively low timeout (or
because some unexpected moving part interferes with the test). There is a time
and place for such a test, but we are currently forced into them because we
often lack a direct way to test at more intermediate levels. This is a problem.

We've seen the benefits of architecting for testability in etcd-io/raft
(applied to *RawNode). A *RawNode has no concurrency, i.e. all access to it
is serial. A set of *RawNodes forms a raft group. The test harness controls
in which order messages get delivered and processed. Verbose logs that result
from handling messages become part of the datadriven output. Everything is
deterministic. The test harness (introduced by us!) combined with the
concurrency-free, deterministic architecture has significantly improved the
ease and quality of testing¹² in etcd/raft.

It stands to reason that the same approach could yield similar benefits if
applied to at least the lower slices in the Replica.Send stack.

Squinting a little (lot), a *Replica that is only accessed in a
single-threaded way is just a *RawNode: it can receive various kinds of
messages (e.g. BatchRequest, raftpb.Message, etc), which it handles
primarily via the raft handling loop; a set of Replicas (for the same range) is
like a Raft group.

Not all of the surface of Replica lends itself to this approach (after all,
some bugs will only surface with concurrency on the Replica, some requests today trigger work on a goroutine, etc) but my intuition and experience
is that with each slice of the stack that we make accessible to the above
strategy will confer a significant improvement of quality.

This broad suggestion leaves much to the imagination and perhaps doesn't make
it clear that in its entirety, this is a significant undertaking. However, small
slices of the stack can be tackled in a bottom-up order and in fact, this is
already happening and has been well-received by all participants:

as part of kvserver: separate raft log #16624, we moved much of the log handling out of the main raft ready loop, into library methods that we could now unit test.
progress was made towards the same goal for log application, and more recently
kvserver: add TestCreateManyUnappliedProbes #104401 made progress for evaluating commands and adding them into the log.

If we could unit test log application, conceivably it would be much easier to
do projects such as #97779 because
we'd be able to directly explore the edge cases in this code, whereas currently
we have to rely on either contrived, hard-to-maintain (and non-exhaustive)
tests, plus randomized tests (which come with their own challenges).

A first "slice" to tackle could be improving the test in
#104401: it creates an
"interesting" raft log programmatically (without even a *Replica), but
currently jumps through some hoops to do so. We could fill these gaps by
successive small refactors, until we have the ability to create arbitrary Raft
logs.

If we then tackled #75729,
i.e. gain the ability to also "apply" these logs in a setting with few moving
parts, we are well underway to being able to comprehensively test the apply
pipeline, which is just not possible today.

We could then work on our ability to instantiate a *Replica "without moving
parts" (i.e. no queues, Raft scheduler, etc) and start writing deterministic
tests.

Jira issue: CRDB-28912

Epic CRDB-39898

The text was updated successfully, but these errors were encountered:

blathers-crl · 2023-06-20T12:52:18Z

cc @cockroachdb/replication

…WithNewLeaseIndexError These tests don't work with v2 and aren't maintainable or super useful anyway. We have good randomized coverage of these code paths now via kvnemesis and TestProposalsWithInjectedLeaseIndexAndReproposalError, and adding better unit test coverage is tracked in cockroachdb#106504. In addition to this, it would be desirable to extend this unit test coverage upwards through the stack via cockroachdb#105177.

This addresses the following race: - n1 runs a ConfChange that adds n2 as a learner. - n1 sends MsgApp to the learner. - n1 starts the INITIAL snapshot, say at index 100. - n2 receives n1's MsgApp. Since it's an uninitialized Replica and its log is empty, it rejects this MsgApp. - n2 receives and applies the INITIAL snapshot, which prompts it to send an affirmative MsgAppResp to n1. - n1's RawNode now tracks n2 as StateProbe (via call to ReportSnapshot(success)) - n1 receives the MsgApp rejection; n2 regresses to StateSnapshot because the rejection comes with a RejectHint (suggested next index to try) of zero, which is not in n1's log. In particular, the SnapshotIndex will likely be higher than the index of the snapshot actually sent, say 101. - n1 receives the affirmative MsgAppResp (for index 100). However, 100 < 101 so this is ignored and the follower remains in StateSnapshot. With this commit, the last two steps cannot happen: n2 transitions straight to StateReplicate because we step a copy of the affirmative MsgAppResp in. The later rejection will be dropped, since it is stale (you can't hint at index zero when you already have a positive index confirmed). I will add that there is no great testing for the above other than stressing the test with additional logging, noting the symptoms, and noting that they disappear with this commit. Scripted testing of this code is within reach[^1] but is outside of the scope of this PR. [^1]: cockroachdb#105177 There is an additional bit of brittleness that is silently suppressed by this commit, but which deserves to be fixed independently because how the problem gets avoided seems accidental and incomplete. When raft requests a snapshot, it notes its current LastIndex and uses it as the PendingSnapshot for the follower's Progress. At the time of writing, MsgAppResp that reconnect the follower to the log but which are not greater than or equal to PendingSnapshot are ignored. In effect, this means that perfectly good snapshots are thrown away if they happen to be a little bit stale. In the example above, the snapshot is stale: PendingSnapshot is 101, but the snapshot is at index 100. Then how does this commit (mostly) fix the problem, i.e. why isn't the snapshot discarded? The key is that when we synchronously step the MsgAppResp(100) into the leader's RawNode, the rejection hasn't arrived yet and so the follower transitions into StateReplicate with a Match of 100. This is then enough so that raft recognizes the rejected MsgApp as stale (since it would regress on durably stored entries). However, there is an alternative example where the rejection arrives earlier: after the snapshot index has been picked, but before the follower has been transitioned into StateReplicate. For this to have a negative effect, an entry has to be appended to the leader's log between generating the snapshot and handling the rejection. Without the combination of delegated snapshots and sustained write activity on the leader, this window is small, and this combination is usually not present in tests but it may well be relevant in "real" clusters. We track addressing this in cockroachdb#106813. Closes cockroachdb#87554. Closes cockroachdb#97971. Closes cockroachdb#84242. Epic: None Release note (bug fix): removed a source of unnecessary Raft snapshots during replica movement.

106793: kvserver: communicate snapshot index back along with snapshot response r=erikgrinaker a=tbg This addresses the following race: - n1 runs a ConfChange that adds n2 as a learner. - n1 sends MsgApp to the learner. - n1 starts the INITIAL snapshot, say at index 100. - n2 receives n1's MsgApp. Since it's an uninitialized Replica and its log is empty, it rejects this MsgApp. - n2 receives and applies the INITIAL snapshot, which prompts it to send an affirmative MsgAppResp to n1. - n1's RawNode now tracks n2 as StateProbe (via call to ReportSnapshot(success)) - n1 receives the MsgApp rejection; n2 regresses to StateSnapshot because the rejection comes with a RejectHint (suggested next index to try) of zero, which is not in n1's log. In particular, the SnapshotIndex will likely be higher than the index of the snapshot actually sent, say 101. - n1 receives the affirmative MsgAppResp (for index 100). However, 100 < 101 so this is ignored and the follower remains in StateSnapshot. With this commit, the last two steps cannot happen: n2 transitions straight to StateReplicate because we step a copy of the affirmative MsgAppResp in. The later rejection will be dropped, since it is stale (you can't hint at index zero when you already have a positive index confirmed). I will add that there is no great testing for the above other than stressing the test with additional logging, noting the symptoms, and noting that they disappear with this commit. Scripted testing of this code is within reach[^1] but is outside of the scope of this PR. [^1]: #105177 There is an additional bit of brittleness that is silently suppressed by this commit, but which deserves to be fixed independently because how the problem gets avoided seems accidental and incomplete. When raft requests a snapshot, it notes its current LastIndex and uses it as the PendingSnapshot for the follower's Progress. At the time of writing, MsgAppResp that reconnect the follower to the log but which are not greater than or equal to PendingSnapshot are ignored. In effect, this means that perfectly good snapshots are thrown away if they happen to be a little bit stale. In the example above, the snapshot is stale: PendingSnapshot is 101, but the snapshot is at index 100. Then how does this commit (mostly) fix the problem, i.e. why isn't the snapshot discarded? The key is that when we synchronously step the MsgAppResp(100) into the leader's RawNode, the rejection hasn't arrived yet and so the follower transitions into StateReplicate with a Match of 100. This is then enough so that raft recognizes the rejected MsgApp as stale (since it would regress on durably stored entries). However, there is an alternative example where the rejection arrives earlier: after the snapshot index has been picked, but before the follower has been transitioned into StateReplicate. For this to have a negative effect, an entry has to be appended to the leader's log between generating the snapshot and handling the rejection. Without the combination of delegated snapshots and sustained write activity on the leader, this window is small, and this combination is usually not present in tests but it may well be relevant in "real" clusters. We track addressing this in #106813. Closes #87554. Closes #97971. Closes #84242. Epic: None Release note (bug fix): removed a source of unnecessary Raft snapshots during replica movement. Co-authored-by: Tobias Grieger <[email protected]> Co-authored-by: Andrew Baptist <[email protected]>

This addresses the following race: - n1 runs a ConfChange that adds n2 as a learner. - n1 sends MsgApp to the learner. - n1 starts the INITIAL snapshot, say at index 100. - n2 receives n1's MsgApp. Since it's an uninitialized Replica and its log is empty, it rejects this MsgApp. - n2 receives and applies the INITIAL snapshot, which prompts it to send an affirmative MsgAppResp to n1. - n1's RawNode now tracks n2 as StateProbe (via call to ReportSnapshot(success)) - n1 receives the MsgApp rejection; n2 regresses to StateSnapshot because the rejection comes with a RejectHint (suggested next index to try) of zero, which is not in n1's log. In particular, the SnapshotIndex will likely be higher than the index of the snapshot actually sent, say 101. - n1 receives the affirmative MsgAppResp (for index 100). However, 100 < 101 so this is ignored and the follower remains in StateSnapshot. With this commit, the last two steps cannot happen: n2 transitions straight to StateReplicate because we step a copy of the affirmative MsgAppResp in. The later rejection will be dropped, since it is stale (you can't hint at index zero when you already have a positive index confirmed). I will add that there is no great testing for the above other than stressing the test with additional logging, noting the symptoms, and noting that they disappear with this commit. Scripted testing of this code is within reach[^1] but is outside of the scope of this PR. [^1]: cockroachdb#105177 There is an additional bit of brittleness that is silently suppressed by this commit, but which deserves to be fixed independently because how the problem gets avoided seems accidental and incomplete. When raft requests a snapshot, it notes its current LastIndex and uses it as the PendingSnapshot for the follower's Progress. At the time of writing, MsgAppResp that reconnect the follower to the log but which are not greater than or equal to PendingSnapshot are ignored. In effect, this means that perfectly good snapshots are thrown away if they happen to be a little bit stale. In the example above, the snapshot is stale: PendingSnapshot is 101, but the snapshot is at index 100. Then how does this commit (mostly) fix the problem, i.e. why isn't the snapshot discarded? The key is that when we synchronously step the MsgAppResp(100) into the leader's RawNode, the rejection hasn't arrived yet and so the follower transitions into StateReplicate with a Match of 100. This is then enough so that raft recognizes the rejected MsgApp as stale (since it would regress on durably stored entries). However, there is an alternative example where the rejection arrives earlier: after the snapshot index has been picked, but before the follower has been transitioned into StateReplicate. For this to have a negative effect, an entry has to be appended to the leader's log between generating the snapshot and handling the rejection. Without the combination of delegated snapshots and sustained write activity on the leader, this window is small, and this combination is usually not present in tests but it may well be relevant in "real" clusters. We track addressing this in cockroachdb#106813. Closes cockroachdb#87554. Closes cockroachdb#97971. Closes cockroachdb#84242. Epic: None Release note (bug fix): removed a source of unnecessary Raft snapshots during replica movement.

…apshot We saw this test hang in CI. What likely happened (according to the stacks) is that a lease transfer that was supposed to be caught by an interceptor never showed up in the interceptor. The most likely explanation is that it errored out before it got to evaluation. It then signaled a channel the test was only prepared to check later, so the test hung (waiting for a channel that was now never to be touched). This test is hard to maintain. It would be great (though, for now, out of reach) to write tests like it in a deterministic framework[^1] [^1]: see cockroachdb#105177. For now, fix the test so that when the (so far unknown) error rears its head again, it will fail properly, so we get to see the error and can take another pass at fixing the test (separately). Stressing this commit[^2], we get: > transferErrC unexpectedly signaled: /Table/Max: transfer lease unexpected > error: refusing to transfer lease to (n3,s3):3 because target may need a Raft > snapshot: replica in StateProbe This makes sense. The test wants to exercise the below-raft mechanism, but the above-raft mechanism also exists and while we didn't want to interact with it, we sometimes do[^1] [^1]: somewhat related to cockroachdb#107524 [^2]: `./dev test --filter TestLeaseTransferRejectedIfTargetNeedsSnapshot --stress ./pkg/kv/kvserver/` on gceworker, 285s Touches cockroachdb#106383. Epic: None Release note: None

107265: liveness: allow registering callbacks after start r=erikgrinaker a=tbg I discovered[^1] a deadlock scenario when multiple nodes in the cluster restart with additional stores that need to be bootstrapped. In that case, liveness must be running when the StoreIDs are allocated, but it is not. Trying to address this problem, I realized that when an auxiliary Store is bootstrapped, it will create a new replicateQueue, which will register a new callback into NodeLiveness. But if liveness must be started at this point to fix #106706, we'll run into the assertion that checks that we don't register callbacks on a started node liveness. Something's got to give: we will allow registering callbacks at any given point in time, and they'll get an initial set of notifications synchronously. I audited the few users of RegisterCallback and this seems OK with all of them. [^1]: #106706 (comment) Epic: None Release note: None 107417: kvserver: ignore RPC conn when deciding to campaign/vote r=erikgrinaker a=erikgrinaker **kvserver: remove stale mayCampaignOnWake comment** The comment is about a parameter that no longer exists. **kvserver: revamp shouldCampaign/Forget tests** **kvserver: ignore RPC conn in `shouldCampaignOnWake`** Previously, `shouldCampaignOnWake()` used `IsLiveMapEntry.IsLive` to determine whether the leader was dead. However, this not only depends on the node's liveness, but also its RPC connectivity. This can prevent an unquiescing replica from acquiring Raft leadership if the leader is still alive but unable to heartbeat liveness, and the leader will be unable to acquire epoch leases in this case. This patch ignores the RPC connection state when deciding whether to campaign, using only on the liveness state. **kvserver: ignore RPC conn in `shouldForgetLeaderOnVoteRequest`** Previously, `shouldForgetLeaderOnVoteRequest()` used `IsLiveMapEntry.IsLive` to determine whether the leader was dead. However, this not only depends on the node's liveness, but also its RPC connectivity. This can prevent granting votes to a new leader that may be attempting to acquire a epoch lease (which the current leader can't). This patch ignores the RPC connection state when deciding whether to campaign, using only on the liveness state. Resolves #107060. Epic: none Release note: None **kvserver: remove `StoreTestingKnobs.DisableLivenessMapConnHealth`** 107424: kvserver: scale Raft entry cache size with system memory r=erikgrinaker a=erikgrinaker The Raft entry cache size defaulted to 16 MB, which is rather small. This has been seen to cause tail latency and throughput degradation with high write volume on large nodes, correlating with a reduction in the entry cache hit rate. This patch linearly scales the Raft entry cache size as 1/256 of total system/cgroup memory, shared evenly between all stores, with a minimum 32 MB. For example, a 32 GB 8-vCPU node will have a 128 MB entry cache. This is a conservative default, since this memory is not accounted for in existing memory budgets nor by the `--cache` flag. We rarely see cache misses in production clusters anyway, and have seen significantly improved hit rates with this scaling (e.g. a 64 KB kv0 workload on 8-vCPU nodes increased from 87% to 99% hit rate). Resolves #98666. Epic: none Release note (performance improvement): The default Raft entry cache size has been increased from 16 MB to 1/256 of system memory with a minimum of 32 MB, divided evenly between all stores. This can be configured via `COCKROACH_RAFT_ENTRY_CACHE_SIZE`. 107442: kvserver: deflake TestRequestsOnFollowerWithNonLiveLeaseholder r=erikgrinaker a=tbg The test previously relied on aggressive liveness heartbeat expirations to avoid running for too long. As a result, it was flaky since liveness wasn't reliably pinned in the way the test wanted. The hybrid manual clock allows time to jump forward at an opportune moment. Use it here to avoid running with a tight lease interval. On my gceworker, previously flaked within a few minutes. As of this commit, I ran it for double-digit minutes without issue. Fixes #107200. Epic: None Release note: None 107526: kvserver: fail gracefully in TestLeaseTransferRejectedIfTargetNeedsSnapshot r=erikgrinaker a=tbg We saw this test hang in CI. What likely happened (according to the stacks) is that a lease transfer that was supposed to be caught by an interceptor never showed up in the interceptor. The most likely explanation is that it errored out before it got to evaluation. It then signaled a channel the test was only prepared to check later, so the test hung (waiting for a channel that was now never to be touched). This test is hard to maintain. It would be great (though, for now, out of reach) to write tests like it in a deterministic framework[^1] [^1]: see #105177. For now, fix the test so that when the (so far unknown) error rears its head again, it will fail properly, so we get to see the error and can take another pass at fixing the test (separately). Stressing this commit[^2], we get: > transferErrC unexpectedly signaled: /Table/Max: transfer lease unexpected > error: refusing to transfer lease to (n3,s3):3 because target may need a Raft > snapshot: replica in StateProbe This makes sense. The test wants to exercise the below-raft mechanism, but the above-raft mechanism also exists and while we didn't want to interact with it, we sometimes do[^1] The second commit introduces a testing knob that disables the above-raft mechanism selectively. I've stressed the test for 15 minutes without issues after this change. [^1]: somewhat related to #107524 [^2]: `./dev test --filter TestLeaseTransferRejectedIfTargetNeedsSnapshot --stress ./pkg/kv/kvserver/` on gceworker, 285s Fixes #106383. Epic: None Release note: None 107531: kvserver: disable replicate queue and lease transfers in closedts tests r=erikgrinaker a=tbg For a more holistic suggestion on how to fix this for the likely many other tests susceptible to similar issues, see: #107528 > 1171 runs so far, 0 failures, over 15m55s Fixes #101824. Release note: None Epic: none Co-authored-by: Tobias Grieger <[email protected]> Co-authored-by: Erik Grinaker <[email protected]>

tbg added C-enhancement Solution expected to add code/behavior + preserve backward-compat (pg compat issues are exception) T-kv-replication labels Jun 20, 2023

tbg mentioned this issue Jun 20, 2023

kvserver: add TestCreateManyUnappliedProbes #104401

Merged

exalate-issue-sync bot added T-kv KV Team and removed T-kv-replication labels Jun 28, 2024

github-project-automation bot added this to KV Aug 28, 2024

github-project-automation bot moved this to Incoming in KV Aug 28, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

kvserver: test Replica like raft.RawNode #105177

kvserver: test Replica like raft.RawNode #105177

tbg commented Jun 20, 2023 •

edited by exalate-issue-sync bot

Loading

blathers-crl bot commented Jun 20, 2023

kvserver: test *Replica like *raft.RawNode #105177

kvserver: test *Replica like *raft.RawNode #105177

Comments

tbg commented Jun 20, 2023 • edited by exalate-issue-sync bot Loading

Footnotes

blathers-crl bot commented Jun 20, 2023

kvserver: test Replica like raft.RawNode #105177

kvserver: test Replica like raft.RawNode #105177

tbg commented Jun 20, 2023 •

edited by exalate-issue-sync bot

Loading