Skip to content
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.

Sign up for GitHub

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

Jump to bottom

roachtest: version/mixed/nodes=3 failed #38560

Closed
cockroach-teamcity opened this issue Jun 28, 2019 · 1 comment · Fixed by #38632
Closed

roachtest: version/mixed/nodes=3 failed #38560

cockroach-teamcity opened this issue Jun 28, 2019 · 1 comment · Fixed by #38632
Assignees
@andreimatei
Labels
C-test-failure Broken test (automatically or manually discovered). O-roachtest O-robot Originated from a bot.
Milestone
19.2

Comments

@cockroach-teamcity
Copy link
Member

SHA: https://github.com/cockroachdb/cockroach/commits/90841a6559df9d9a4724e1d30490951bbdb811b4

Parameters:

To repro, try:

# Don't forget to check out a clean suitable branch and experiment with the
# stress invocation until the desired results present themselves. For example,
# using stress instead of stressrace and passing the '-p' stressflag which
# controls concurrency.
./scripts/gceworker.sh start && ./scripts/gceworker.sh mosh
cd ~/go/src/github.com/cockroachdb/cockroach && \
stdbuf -oL -eL \
make stressrace TESTS=version/mixed/nodes=3 PKG=roachtest TESTTIMEOUT=5m STRESSFLAGS='-maxtime 20m -timeout 10m' 2>&1 | tee /tmp/stress.log

Failed test: https://teamcity.cockroachdb.com/viewLog.html?buildId=1364443&tab=buildLog

The test failed on branch=provisional_201906271846_v19.2.0-alpha.20190701, cloud=gce:
	cluster.go:1870,version.go:233,version.go:246,test.go:1249: signal: interrupt
	cluster.go:1587,cluster.go:1606,cluster.go:1710,cluster.go:1093,context.go:122,cluster.go:1090,panic.go:406,test.go:783,test.go:769,cluster.go:1870,version.go:233,version.go:246,test.go:1249: context canceled
@cockroach-teamcity cockroach-teamcity added this to the 19.2 milestone Jun 28, 2019
@cockroach-teamcity cockroach-teamcity added C-test-failure Broken test (automatically or manually discovered). O-roachtest O-robot Originated from a bot. labels Jun 28, 2019
@cockroach-teamcity
Copy link
Member Author

SHA: https://github.com/cockroachdb/cockroach/commits/537767ac9daa52b0026bb957d7010e3b88b61071

Parameters:

To repro, try:

# Don't forget to check out a clean suitable branch and experiment with the
# stress invocation until the desired results present themselves. For example,
# using stress instead of stressrace and passing the '-p' stressflag which
# controls concurrency.
./scripts/gceworker.sh start && ./scripts/gceworker.sh mosh
cd ~/go/src/github.com/cockroachdb/cockroach && \
stdbuf -oL -eL \
make stressrace TESTS=version/mixed/nodes=3 PKG=roachtest TESTTIMEOUT=5m STRESSFLAGS='-maxtime 20m -timeout 10m' 2>&1 | tee /tmp/stress.log

Failed test: https://teamcity.cockroachdb.com/viewLog.html?buildId=1364821&tab=buildLog

The test failed on branch=master, cloud=gce:
	cluster.go:1870,version.go:233,version.go:246,test.go:1249: signal: interrupt
	cluster.go:1587,cluster.go:1606,cluster.go:1710,cluster.go:1093,context.go:122,cluster.go:1090,panic.go:406,test.go:783,test.go:769,cluster.go:1870,version.go:233,version.go:246,test.go:1249: context canceled

nvanbenschoten added a commit to nvanbenschoten/cockroach that referenced this issue Jul 3, 2019
@nvanbenschoten
storage: release quota on failed Raft proposals
23b1fb9 
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 for 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
@nvanbenschoten nvanbenschoten mentioned this issue Jul 3, 2019
Merged
nvanbenschoten added a commit to nvanbenschoten/cockroach that referenced this issue Jul 3, 2019
@nvanbenschoten
storage: release quota on failed Raft proposals
ba3813c 
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
craig bot pushed a commit that referenced this issue Jul 3, 2019
@nvanbenschoten
Merge #38632
40aac81 
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>
@craig craig bot closed this as completed in #38632 Jul 3, 2019
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees

@andreimatei andreimatei

Labels
C-test-failure Broken test (automatically or manually discovered). O-roachtest O-robot Originated from a bot.
Projects
None yet
Milestone
19.2
Development

Successfully merging a pull request may close this issue.

storage: release quota on failed Raft proposals nvanbenschoten/cockroach
2 participants
@andreimatei @cockroach-teamcity