Computing the snapshot file path is error prone; snapshot recovery was
constructing file paths missing a path separator so the snapshot
would never be loaded. Instead, refactor the backend path handling
to use helper functions where possible.
In the case that follower recieves a snapshot from leader
and crashes before renaming xxx.snap.db to db but after
snapshot has persisted to .wal and .snap, restarting
follower results loading old db, new .wal, and new .snap.
This will causes a index mismatch between snap metadata index
and consistent index from db.
This pr forces an ordering where saving/renaming db must
happen after snapshot is persisted to wal and snap file.
this guarantees wal and snap files are newer than db.
on server restart, etcd server checks if snap index > db consistent index.
if yes, etcd server attempts to load xxx.snap.db where xxx=snap index
if there is any and panic other wise.
FIXES#7628
previously, apply() doesn't set consistIndex for EntryConfChange type.
this causes a misalignment between consistIndex and applied index
where EntryConfChange entry results setting applied index but not consistIndex.
suppose that addMember() is called and leader reflects that change.
1. applied index and consistIndex is now misaligned.
2. a new follower node joined.
3. leader sends the snapshot to follower
where the applied index is the snapshot metadata index.
4. follower node saves the snapshot and database(includes consistIndex) from leader.
5. restarting follower loads snapshot and database.
6. follower checks snapshot metadata index(same as applied index) and database consistIndex,
finds them don't match, and then panic.
FIXES#7834
Printing the values in ctx.String() will data race if the value
is mutable and doesn't implement String(), which seems to be common.
Instead, just return a fixed string instead of computing it; v3client
watches don't need as much flexibility for creating separate strings,
so separate ctx strings probably aren't necessary at this point.
Fixes#7811
Problem is:
`Step1`: `etcdserver/raft.go`'s `Ready` process routine sends config-change entries via `r.applyc <- ap` (https://github.com/coreos/etcd/blob/master/etcdserver/raft.go#L193-L203)
`Step2`: `etcdserver/server.go`'s `*EtcdServer.run` routine receives this via `ap := <-s.r.apply()` (https://github.com/coreos/etcd/blob/master/etcdserver/server.go#L735-L738)
`StepA`: `Step1` proceeds without sync, right after sending `r.applyc <- ap`.
`StepB`: `Step2` proceeds without sync, right after `sched.Schedule(s.applyAll(&ep,&ap))`.
`StepC`: `etcdserver` tries to sync with `s.applyAll(&ep,&ap)` by calling `rh.waitForApply()`.
`rh.waitForApply()` waits for all pending jobs to finish in `pkg/schedule`
side. However, the order of `StepA`,`StepB`,`StepC` is not guaranteed. It
is possible that `StepC` happens first, and proceeds without waiting on
apply. And the restarting member comes back as a leader in single-node
cluster, when there is no synchronization between apply-layer and
config-change Raft entry apply. Confirmed with more debugging lines below,
only reproducible with slow CPU VM (~2 vCPU).
```
~:24.005397 I | etcdserver: starting server... [version: 3.2.0+git, cluster version: to_be_decided]
~:24.011136 I | etcdserver: [DEBUG] 29b2d24047a277df waitForApply before
~:24.011194 I | etcdserver: [DEBUG] 29b2d24047a277df starts wait for 0 pending jobs
~:24.011234 I | etcdserver: [DEBUG] 29b2d24047a277df finished wait for 0 pending jobs (current pending 0)
~:24.011268 I | etcdserver: [DEBUG] 29b2d24047a277df waitForApply after
~:24.011348 I | etcdserver: [DEBUG] [0] 29b2d24047a277df is scheduling conf change on 29b2d24047a277df
~:24.011396 I | etcdserver: [DEBUG] [1] 29b2d24047a277df is scheduling conf change on 5edf80e32a334cf0
~:24.011437 I | etcdserver: [DEBUG] [2] 29b2d24047a277df is scheduling conf change on e32e31e76c8d2678
~:24.011477 I | etcdserver: [DEBUG] 29b2d24047a277df scheduled conf change on 29b2d24047a277df
~:24.011509 I | etcdserver: [DEBUG] 29b2d24047a277df scheduled conf change on 5edf80e32a334cf0
~:24.011545 I | etcdserver: [DEBUG] 29b2d24047a277df scheduled conf change on e32e31e76c8d2678
~:24.012500 I | etcdserver: [DEBUG] 29b2d24047a277df applyConfChange on 29b2d24047a277df before
~:24.013014 I | etcdserver/membership: added member 29b2d24047a277df [unix://127.0.0.1:2100515039] to cluster 9250d4ae34216949
~:24.013066 I | etcdserver: [DEBUG] 29b2d24047a277df applyConfChange on 29b2d24047a277df after
~:24.013113 I | etcdserver: [DEBUG] 29b2d24047a277df applyConfChange on 29b2d24047a277df after trigger
~:24.013158 I | etcdserver: [DEBUG] 29b2d24047a277df applyConfChange on 5edf80e32a334cf0 before
~:24.013666 W | etcdserver: failed to send out heartbeat on time (exceeded the 10ms timeout for 11.964739ms)
~:24.013709 W | etcdserver: server is likely overloaded
~:24.013750 W | etcdserver: failed to send out heartbeat on time (exceeded the 10ms timeout for 12.057265ms)
~:24.013775 W | etcdserver: server is likely overloaded
~:24.013950 I | raft: 29b2d24047a277df is starting a new election at term 4
~:24.014012 I | raft: 29b2d24047a277df became candidate at term 5
~:24.014051 I | raft: 29b2d24047a277df received MsgVoteResp from 29b2d24047a277df at term 5
~:24.014107 I | raft: 29b2d24047a277df became leader at term 5
~:24.014146 I | raft: raft.node: 29b2d24047a277df elected leader 29b2d24047a277df at term 5
```
I am printing out the number of pending jobs before we call
`sched.WaitFinish(0)`, and there was no pending jobs, so it returned
immediately (before we schedule `applyAll`).
This is the root cause to:
- https://github.com/coreos/etcd/issues/7595
- https://github.com/coreos/etcd/issues/7739
- https://github.com/coreos/etcd/issues/7802
`sched.WaitFinish(0)` doesn't work when `len(f.pendings)==0` and
`f.finished==0`. Config-change is the first job to apply, so
`f.finished` is 0 in this case.
`f.finished` monotonically increases, so we need `WaitFinish(finished+1)`.
And `finished` must be the one before calling `Schedule`. This is safe
because `Schedule(applyAll)` is the only place adding jobs to `sched`.
Then scheduler waits on the single job of `applyAll`, by getting the
current number of finished jobs before sending `Schedule`.
Or just make it be blocked until `applyAll` routine triggers on the
config-change job. This patch just removes `waitForApply`, and
signal `raftDone` to wait until `applyAll` finishes applying entries.
Confirmed that it fixes the issue, as below:
```
~:43.198354 I | rafthttp: started streaming with peer 36cda5222aba364b (stream MsgApp v2 reader)
~:43.198740 I | etcdserver: [DEBUG] 3988bc20c2b2e40c waitForApply before
~:43.198836 I | etcdserver: [DEBUG] 3988bc20c2b2e40c starts wait for 0 pending jobs, 1 finished jobs
~:43.200696 I | integration: launched 3169361310155633349 ()
~:43.201784 I | etcdserver: [DEBUG] [0] 3988bc20c2b2e40c is scheduling conf change on 36cda5222aba364b
~:43.201884 I | etcdserver: [DEBUG] [1] 3988bc20c2b2e40c is scheduling conf change on 3988bc20c2b2e40c
~:43.201965 I | etcdserver: [DEBUG] [2] 3988bc20c2b2e40c is scheduling conf change on cf5d6cbc2a121727
~:43.202070 I | etcdserver: [DEBUG] 3988bc20c2b2e40c scheduled conf change on 36cda5222aba364b
~:43.202139 I | etcdserver: [DEBUG] 3988bc20c2b2e40c scheduled conf change on 3988bc20c2b2e40c
~:43.202204 I | etcdserver: [DEBUG] 3988bc20c2b2e40c scheduled conf change on cf5d6cbc2a121727
~:43.202444 I | etcdserver: [DEBUG] 3988bc20c2b2e40c applyConfChange on 36cda5222aba364b (request ID: 0) before
~:43.204486 I | etcdserver/membership: added member 36cda5222aba364b [unix://127.0.0.1:2100913646] to cluster 425d73f1b7b01674
~:43.204588 I | etcdserver: [DEBUG] 3988bc20c2b2e40c applyConfChange on 36cda5222aba364b (request ID: 0) after
~:43.204703 I | etcdserver: [DEBUG] 3988bc20c2b2e40c applyConfChange on 36cda5222aba364b (request ID: 0) after trigger
~:43.204791 I | etcdserver: [DEBUG] 3988bc20c2b2e40c applyConfChange on 3988bc20c2b2e40c (request ID: 0) before
~:43.205689 I | etcdserver/membership: added member 3988bc20c2b2e40c [unix://127.0.0.1:2101113646] to cluster 425d73f1b7b01674
~:43.205783 I | etcdserver: [DEBUG] 3988bc20c2b2e40c applyConfChange on 3988bc20c2b2e40c (request ID: 0) after
~:43.205929 I | etcdserver: [DEBUG] 3988bc20c2b2e40c applyConfChange on 3988bc20c2b2e40c (request ID: 0) after trigger
~:43.206056 I | etcdserver: [DEBUG] 3988bc20c2b2e40c applyConfChange on cf5d6cbc2a121727 (request ID: 0) before
~:43.207353 I | etcdserver/membership: added member cf5d6cbc2a121727 [unix://127.0.0.1:2100713646] to cluster 425d73f1b7b01674
~:43.207516 I | etcdserver: [DEBUG] 3988bc20c2b2e40c applyConfChange on cf5d6cbc2a121727 (request ID: 0) after
~:43.207619 I | etcdserver: [DEBUG] 3988bc20c2b2e40c applyConfChange on cf5d6cbc2a121727 (request ID: 0) after trigger
~:43.207710 I | etcdserver: [DEBUG] 3988bc20c2b2e40c finished scheduled conf change on 36cda5222aba364b
~:43.207781 I | etcdserver: [DEBUG] 3988bc20c2b2e40c finished scheduled conf change on 3988bc20c2b2e40c
~:43.207843 I | etcdserver: [DEBUG] 3988bc20c2b2e40c finished scheduled conf change on cf5d6cbc2a121727
~:43.207951 I | etcdserver: [DEBUG] 3988bc20c2b2e40c finished wait for 0 pending jobs (current pending 0, finished 1)
~:43.208029 I | rafthttp: started HTTP pipelining with peer cf5d6cbc2a121727
~:43.210339 I | rafthttp: peer 3988bc20c2b2e40c became active
~:43.210435 I | rafthttp: established a TCP streaming connection with peer 3988bc20c2b2e40c (stream MsgApp v2 reader)
~:43.210861 I | rafthttp: started streaming with peer 3988bc20c2b2e40c (writer)
~:43.211732 I | etcdserver: [DEBUG] 3988bc20c2b2e40c waitForApply after
```
Signed-off-by: Gyu-Ho Lee <gyuhox@gmail.com>
raftNode was being initialized in start(), which was causing
hangs when trying to stop the etcd server since the stop channel
would not be initialized in time for the stop call. Instead,
setup non-configurable bits in a constructor.
Fixes#7668
Replacing "etcdserver: fill a response header in auth RPCs"
The revision should be set at the time of "apply",
not in later RPC layer.
Fix https://github.com/coreos/etcd/issues/7691
Signed-off-by: Gyu-Ho Lee <gyuhox@gmail.com>
When apply-layer sees configuration change entry in
raft.Ready.CommittedEntries, the server should not proceed
until that entry is applied. Otherwise, follower's raft
layer advances, possibly election-timeouts, and becomes
the leader in single-node cluster, before add-node conf
change of other nodes is applied.
Fix https://github.com/coreos/etcd/issues/7595.
Signed-off-by: Gyu-Ho Lee <gyuhox@gmail.com>
gRPC will replace empty strings with nil, but for the embedded case it's
possible for []byte{} to slip in and confuse the single key / >= key
watch logic.
Fix https://github.com/coreos/etcd/issues/7470.
This patch removes unnecessary term look-up in
'createMergedSnapshotMessage', which can trigger panic
if raft entry at etcdProgress.appliedi got compacted
by subsequent 'MsgSnap' messages--if a follower is
being (in this case, network latency spikes), it
could receive subsequent 'MsgSnap' requests from leader.
etcd server-side 'applyAll' routine and raft's Ready
processing routine becomes asynchronous after raft
entries are persisted. And given that raft Ready routine
takes less time to finish, it is possible that second
'MsgSnap' is being handled, while the slow 'applyAll'
is still processing the first(old) 'MsgSnap'. Then raft
Ready routine can compact the log entries at future
index to 'applyAll'. That is how 'createMergedSnapshotMessage'
tried to look up raft term with outdated etcdProgress.appliedi.
Signed-off-by: Gyu-Ho Lee <gyuhox@gmail.com>
This commit adds jwt token support in v3 auth API.
Remaining major ToDos:
- Currently token type isn't hidden from etcdserver. In the near
future the information should be completely invisible from
etcdserver package.
- Configurable expiration of token. Currently tokens can be valid
until keys are changed.
How to use:
1. generate keys for signing and verfying jwt tokens:
$ openssl genrsa -out app.rsa 1024
$ openssl rsa -in app.rsa -pubout > app.rsa.pub
2. add command line options to etcd like below:
--auth-token-type jwt \
--auth-jwt-pub-key app.rsa.pub --auth-jwt-priv-key app.rsa \
--auth-jwt-sign-method RS512
3. launch etcd cluster
Below is a performance comparison of serializable read w/ and w/o jwt
token. Every (3) etcd node is executed on a single machine. Signing
method is RS512 and key length is 1024 bit. As the results show, jwt
based token introduces a performance overhead but it would be
acceptable for a case that requires authentication.
w/o jwt token auth (no auth):
Summary:
Total: 1.6172 secs.
Slowest: 0.0125 secs.
Fastest: 0.0001 secs.
Average: 0.0002 secs.
Stddev: 0.0004 secs.
Requests/sec: 6183.5877
Response time histogram:
0.000 [1] |
0.001 [9982] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
0.003 [1] |
0.004 [1] |
0.005 [0] |
0.006 [0] |
0.008 [6] |
0.009 [0] |
0.010 [1] |
0.011 [5] |
0.013 [3] |
Latency distribution:
10% in 0.0001 secs.
25% in 0.0001 secs.
50% in 0.0001 secs.
75% in 0.0001 secs.
90% in 0.0002 secs.
95% in 0.0002 secs.
99% in 0.0003 secs.
w/ jwt token auth:
Summary:
Total: 2.5364 secs.
Slowest: 0.0182 secs.
Fastest: 0.0002 secs.
Average: 0.0003 secs.
Stddev: 0.0005 secs.
Requests/sec: 3942.5185
Response time histogram:
0.000 [1] |
0.002 [9975] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
0.004 [0] |
0.006 [1] |
0.007 [11] |
0.009 [2] |
0.011 [4] |
0.013 [5] |
0.015 [0] |
0.016 [0] |
0.018 [1] |
Latency distribution:
10% in 0.0002 secs.
25% in 0.0002 secs.
50% in 0.0002 secs.
75% in 0.0002 secs.
90% in 0.0003 secs.
95% in 0.0003 secs.
99% in 0.0004 secs.
Getting gosimple suggestion while running test script, so this PR is for fixing gosimple S1019 check.
raft/node_test.go:456:40: should use make([]raftpb.Entry, 1) instead (S1019)
raft/node_test.go:457:49: should use make([]raftpb.Entry, 1) instead (S1019)
raft/node_test.go:458:43: should use make([]raftpb.Message, 1) instead (S1019)
Refer https://github.com/dominikh/go-tools/blob/master/cmd/gosimple/README.md#checks for more information.
Keep more wal entries in memory for fast follower recovery.
10,000 was a too small number that triggers quite a few snapshots.
ZK proves that 100,000 is a reasonable number for even old less prowerful
machines.
Eventually we should provide both count and max memory (for large entries).
This commit adds a mechanism of handling a case of expired auth token
to clientv3. If a server returns an error code
grpc.codes.Unauthenticated, newRetryWrapper() tries to get a new token
and use it as an option of PerRPCCredential.
Fixes https://github.com/coreos/etcd/issues/7012
GetUser would not propagate to the minority node, causing TestCtlV2GetRoleUser to
run CreateUser instead of UpdateUser. Instead, use quorum get to fetch the
current state of auth.
Fixes#7069
Removing the periodic SYNC calls broke the health endpoint since the
raft index stops updating. Instead, don't bother monitoring the
raft index; issue a QGET directly to get a consensus response.
Fixes#6985
suppose a lease granting request from a follower goes through and followed by a lease look up or renewal, the leader might not apply the lease grant request locally. So the leader might not find the lease from the lease look up or renewal request which will result lease not found error. To fix this issue, we force the leader to apply its pending commited index before looking up lease.
FIX#6978
This commit protects membership change operations with auth. Only
users that have root role can issue the operations.
Implements https://github.com/coreos/etcd/issues/6899
Would retry a few times before returning a not primary error that
the client should never see. Instead, use proper timeouts and
then return a request timeout error on failure.
Fixes#6922
This exists to prevent sending too many requests that
would lead into applier falling behind Raft accepting-proposal.
Based on recent benchmarks, etcd was able to process high workloads
(2 million writes with 1K concurrent clients).
The limit 1000 is too conservative to test those high workloads.
Address https://github.com/coreos/etcd/issues/6754.
In case there are network errors or unexpected EOF errors
in TimeToLive http requests to leader, we translate that into
ErrNoLeader, and expects the client to retry its request.
The cost of bcrypt password checking is quite high (almost 100ms on a
modern machine) so executing it in apply loop will be
problematic. This commit exclude the checking mechanism to the API
layer. The password checking is validated with the OCC like way
similar to the auth of serializable get.
This commit also removes a unit test of Authenticate RPC from
auth/store_test.go. It is because the RPC now accepts an auth request
unconditionally and delegates the checking functionality to
authStore.CheckPassword() (so a unit test for CheckPassword() is
added). The combination of the two functionalities can be tested by
e2e (e.g. TestCtlV3AuthWriteKey).
Fixes https://github.com/coreos/etcd/issues/6530
When running test suites for a client locally I'm getting spammed by log
lines such as:
etcdserver: apply entries took too long [14.226771ms for 1 entries]
The comments in #6278 mention there were future plans of increasing the
threshold for logging these warnings, but it hadn't been done yet.
If we promote the lessor before finish applying all
entries from the last term, we might incorrectly renew
the already revoked leases.
Here is an example:
- Term 1: revoke lease A accepted by raft
- Old leader failed, new election happened
- Term 2: promote
- Term 2: keep alive A succeed. A now has 10 seconds TTL
- Term 2: revoke lease A from Term 1 got committed and applied
- Term 2: the lease A with 10 seconds TTL is revoked
To solve this, the new leader MUST apply all entries from old term
before promote its lessor to start accept renew requests.
When the non Leader etcd server receives a LeaseTimeToLive on a nonexistent lease, it responds with a nil resp and a nil error The invoking function parses the nil resp and results a segmentation fault.
I fix the bug by making sure the lease not found error is returned so that the invoking function parses the the error message instead.
fix#6537
When 1000 leases expired at the same time, etcd takes more than 5 seconds to clean them. This means that even after the leases have expired, keys associated with leases are still accessible. I increase the deletion throughput by parallelizing leases deletion process.
All outstanding goroutines now go into the etcdserver waitgroup. goroutines are
shutdown with a "stopping" channel which is closed when the run() goroutine
shutsdown. The done channel will only close once the waitgroup is totally cleared.
Added support into the v2 API to fix an issue (6433) where if there is a semicolon
and fields after it the API would return an "invalid Content-type" message even
if the content type was actually correct
If a user upgrades etcd from 2.3.x to 3.0 and shutdown the
cluster immediately without triggering any new backend writes,
then the consistent index in backend would be zero.
The user cannot restart etcdserver due to today's strick index
match checking. We now have to lose this a bit for this case.
kv.commit updates the consistent index in backend. When
executing in parallel with apply, it might grab tx lock
after apply update the consistent index and before apply
starts to execute the opeartion. If the server dies right
after kv.commit, the consistent is updated but the opeartion
is not executed. If we restart etcd server, etcd will skip
the operation. :(
There are a few other places that we need to take care of,
but let us fix this first.
moved the code for perparing and sorting of advertising peer urls and
sorting of peer urls only when strict verification needs to be done.
This is done to avoid this processing when strict verification is not
required like in case of VerifyJoinExisting function.
#6165
If a server isn't serving txn requests from a client, the server
doesn't need the result of range requests in the txn.
This is a succeeding commit of
https://github.com/coreos/etcd/pull/5689
This commit introduces revision of authStore. The revision number
represents a version of authStore that is incremented by updating auth
related information.
The revision is required for avoiding TOCTOU problems. Currently there
are two types of the TOCTOU problems in v3 auth.
The first one is in ordinal linearizable requests with a sequence like
below ():
1. Request from client CA is processed in follower FA. FA looks up the
username (let it U) for the request from a token of the request. At
this time, the request is authorized correctly.
2. Another request from client CB is processed in follower FB. CB
is for changing U's password.
3. FB forwards the request from CB to the leader before FA. Now U's
password is updated and the request from CA should be rejected.
4. However, the request from CA is processed by the leader because
authentication is already done in FA.
For avoiding the above sequence, this commit lets
etcdserverpb.RequestHeader have a member revision. The member is
initialized during authentication by followers and checked in a
leader. If the revision in RequestHeader is lower than the leader's
authStore revision, it means a sequence like above happened. In such a
case, the state machine returns auth.ErrAuthRevisionObsolete. The
error code lets nodes retry their requests.
The second one, a case of serializable range and txn, is more
subtle. Because these requests are processed in follower directly. The
TOCTOU problem can be caused by a sequence like below:
1. Serializable request from client CA is processed in follower FA. At
first, FA looks up the username (let it U) and its permission
before actual access to KV.
2. Another request from client CB is processed in follower FB and
forwarded to the leader. The cluster including FA now commits a log
entry of the request from CB. Assume the request changed the
permission or password of U.
3. Now the serializable request from CA is accessing to KV. Even if
the access is allowed at the point of 1, now it can be invalid
because of the change introduced in 2.
For avoiding the above sequence, this commit lets the functions of
serializable requests (EtcdServer.Range() and EtcdServer.Txn())
compare the revision in the request header with the latest revision of
authStore after the actual access. If the saved revision is lower than
the latest one, it means the permission can be changed. Although it
would introduce false positives (e.g. changing other user's password),
it prevents the TOCTOU problem. This idea is an implementation of
Anthony's comment:
https://github.com/coreos/etcd/pull/5739#issuecomment-228128254
Most fields accessed with sync/atomic functions are 64bit aligned, but a couple
are not. This makes comments out of date and therefore misleading.
Affected fields reordered, comments scrubbed and updated.
This commit lets etcdserver skip needless log entry applying. If the
result of log applying isn't required by the node (client that issued
the request isn't talking with the node) and the operation has no side
effects, applying can be skipped.
It would contribute to reduce disk I/O on followers and be useful for
a cluster that processes much serializable get.
Before
2016-07-01 14:57:50.927170 I | api: enabled capabilities for version 3.0.0
After
2016-07-01 14:57:50.927170 I | api: enabled capabilities for version 3.0
Now user can filter events with types. The API is also extensible.
It might make sense for the proxy to filter out events based on
more expensive/customized filter.