-
Notifications
You must be signed in to change notification settings - Fork 719
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
server: use leader lease to determine tso service validity #1676
Conversation
Signed-off-by: disksing <i@disksing.com>
Signed-off-by: disksing <i@disksing.com>
Codecov Report
@@ Coverage Diff @@
## master #1676 +/- ##
==========================================
- Coverage 76.3% 76.26% -0.04%
==========================================
Files 157 158 +1
Lines 15347 15402 +55
==========================================
+ Hits 11710 11747 +37
- Misses 2644 2654 +10
- Partials 993 1001 +8
Continue to review full report at Codecov.
|
server/lease.go
Outdated
if cost := time.Since(start); cost > slowRequestTime { | ||
log.Warn("lease grants too slow", zap.Duration("cost", cost)) | ||
} | ||
if err != nil { |
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.
How about returning this error before line 51?
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.
Good idea.
server/lease.go
Outdated
func (l *LeaderLease) KeepAlive(ctx context.Context) { | ||
ctx, cancel := context.WithCancel(ctx) | ||
defer cancel() | ||
timeCh := l.keepAliveWorker(ctx, l.leaseTimeout/3) |
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.
Why use 3
here?
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.
Arbitrary value borrowed from etcd's keep alive code.
server/lease.go
Outdated
|
||
// Close releases the lease. | ||
func (l *LeaderLease) Close() error { | ||
return l.lease.Close() |
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.
May we need Revoke
the lease
before close? actually Close
do not try to release the lease.
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.
LGTM
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.
LGTM
Signed-off-by: disksing <i@disksing.com>
Signed-off-by: disksing <i@disksing.com>
Signed-off-by: disksing <i@disksing.com>
/test |
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.
LGTM
/build |
Created an issue for the test failure: #1693 |
* *: unify get store function everywhere (#1671) Signed-off-by: Ryan Leung <rleungx@gmail.com> * server: use leader lease to determine tso service validity (#1676) Signed-off-by: disksing <i@disksing.com> * test: fix tests (#1696) * test: fix region syncer test Signed-off-by: disksing <i@disksing.com> * add config-check flag for pd-server (#1695) Signed-off-by: cwen0 <cwenyin0@gmail.com> * operator: rewrite move region related functions (#1667) * *: support setting endKey for ScanRange (#1700) Signed-off-by: disksing <i@disksing.com> * *: reduce some unnecessary parameters (#1698) Signed-off-by: Ryan Leung <rleungx@gmail.com> * schedule: Do not send an operator of a region wth a stale epoch (#1659) * schedule: Do not send an operator of a region wth a stale epoch Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * schedule: check the version changed by the operator self Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * schedule: fix unit test Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * schedule: fix to avoid dispatching a stale opstep Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * dispatch: refactor "ConsumeConfVer() int" to "ExpectConfVerChange() bool" Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * dispatch: fix typo in comment Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * fix typo Co-Authored-By: Ryan Leung <rleungx@gmail.com> * dispatch: fix unittest Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * dispatch: refine format Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * server: fix the dead lock in scatter region (#1706) Signed-off-by: Ryan Leung <rleungx@gmail.com>
* *: unify get store function everywhere (#1671) Signed-off-by: Ryan Leung <rleungx@gmail.com> * remove unnecessary parentheses * server: use leader lease to determine tso service validity (#1676) Signed-off-by: disksing <i@disksing.com> * change internal stat values to float64 * add pending operator influence * add metrics of pending influence * fix metrics * fix panic * adjust pending influence of balanceHotWrite * change weight of pending influence * test: fix tests (#1696) * test: fix region syncer test Signed-off-by: disksing <i@disksing.com> * decrease region rolling window; store pending influence in scheduler * add config-check flag for pd-server (#1695) Signed-off-by: cwen0 <cwenyin0@gmail.com> * decrease possiblility transfer hot write leader * change pending influence weight * add unstarted op metrics * add logs for debug * add log for debug * add logs for debug * add logs for debug * add logs for debug * add logs for debug * add logs for debug * add logs for debug * Revert "add logs for debug" This reverts commit e74c7a9. * add metrics for hotspot operators * operator: rewrite move region related functions (#1667) * add metrics for pending operators * *: support setting endKey for ScanRange (#1700) Signed-off-by: disksing <i@disksing.com> * fix bug * fix bug * fix bug * fix metrics thread-safe bug * fix logic bug * *: reduce some unnecessary parameters (#1698) Signed-off-by: Ryan Leung <rleungx@gmail.com> * schedule: Do not send an operator of a region wth a stale epoch (#1659) * schedule: Do not send an operator of a region wth a stale epoch Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * schedule: check the version changed by the operator self Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * schedule: fix unit test Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * schedule: fix to avoid dispatching a stale opstep Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * dispatch: refactor "ConsumeConfVer() int" to "ExpectConfVerChange() bool" Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * dispatch: fix typo in comment Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * fix typo Co-Authored-By: Ryan Leung <rleungx@gmail.com> * dispatch: fix unittest Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * dispatch: refine format Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * server: fix the dead lock in scatter region (#1706) Signed-off-by: Ryan Leung <rleungx@gmail.com> * add drop time for operator * use IsDropped to recognize canceled ops * try to fix trans leader burst * try to fix trans leader burst * add zombie influence * change select src dst strategy; improve op_controller * change select src strategy * fix bug * tools: fix set namespace in pd-ctl (#1701) Signed-off-by: Ryan Leung <rleungx@gmail.com> * tools: fix parse url without http prefix (#1703) Signed-off-by: Ryan Leung <rleungx@gmail.com> * tests: support deadlock detection in make test (#1704) Signed-off-by: Ryan Leung <rleungx@gmail.com> * Makefile: fix failpoint enable (#1722) Signed-off-by: nolouch <nolouch@gmail.com> * checker: fix the issue that a region does not merge to the sibling with smaller size (#1723) Signed-off-by: disksing <i@disksing.com> * tools: balance region simulator (#1708) * scheduler: do not remove the operator when the step does not finish (#1715) Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * operator: fix the AddLearner config version judgment (#1732) Signed-off-by: nolouch <nolouch@gmail.com> * tools: fix TLS in pd control (#1729) Signed-off-by: Ryan Leung <rleungx@gmail.com> * syncer: support TLS for region syncer (#1728) Signed-off-by: Ryan Leung <rleungx@gmail.com> * schedule: fix a thread-safe bug and improve code (#1719)
@disksing should we cherry pick to 3.0,3.1? |
* *: unify get store function everywhere (#1671) Signed-off-by: Ryan Leung <rleungx@gmail.com> * server: use leader lease to determine tso service validity (#1676) Signed-off-by: disksing <i@disksing.com> * test: fix tests (#1696) * test: fix region syncer test Signed-off-by: disksing <i@disksing.com> * add config-check flag for pd-server (#1695) Signed-off-by: cwen0 <cwenyin0@gmail.com> * operator: rewrite move region related functions (#1667) * *: support setting endKey for ScanRange (#1700) Signed-off-by: disksing <i@disksing.com> * *: reduce some unnecessary parameters (#1698) Signed-off-by: Ryan Leung <rleungx@gmail.com> * schedule: Do not send an operator of a region wth a stale epoch (#1659) * schedule: Do not send an operator of a region wth a stale epoch Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * schedule: check the version changed by the operator self Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * schedule: fix unit test Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * schedule: fix to avoid dispatching a stale opstep Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * dispatch: refactor "ConsumeConfVer() int" to "ExpectConfVerChange() bool" Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * dispatch: fix typo in comment Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * fix typo Co-Authored-By: Ryan Leung <rleungx@gmail.com> * dispatch: fix unittest Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * dispatch: refine format Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * server: fix the dead lock in scatter region (#1706) Signed-off-by: Ryan Leung <rleungx@gmail.com> * tools: fix set namespace in pd-ctl (#1701) Signed-off-by: Ryan Leung <rleungx@gmail.com> * tools: fix parse url without http prefix (#1703) Signed-off-by: Ryan Leung <rleungx@gmail.com> * tests: support deadlock detection in make test (#1704) Signed-off-by: Ryan Leung <rleungx@gmail.com> * Makefile: fix failpoint enable (#1722) Signed-off-by: nolouch <nolouch@gmail.com> * checker: fix the issue that a region does not merge to the sibling with smaller size (#1723) Signed-off-by: disksing <i@disksing.com> * tools: balance region simulator (#1708) * scheduler: do not remove the operator when the step does not finish (#1715) Signed-off-by: Shafreeck Sea <shafreeck@gmail.com> * operator: fix the AddLearner config version judgment (#1732) Signed-off-by: nolouch <nolouch@gmail.com> * tools: fix TLS in pd control (#1729) Signed-off-by: Ryan Leung <rleungx@gmail.com> * syncer: support TLS for region syncer (#1728) Signed-off-by: Ryan Leung <rleungx@gmail.com> * schedule: fix a thread-safe bug and improve code (#1719) * statistics: fix region flow calculation (#1688) Signed-off-by: jiyingtk <jiyingtk@mail.ustc.edu.cn> * makefile: improve deadlock-enable/disable (#1736) * api: fix missing keys statistic in region information (#1741) Signed-off-by: nolouch <nolouch@gmail.com> * *: update go version to 1.13 (#1742) Signed-off-by: disksing <i@disksing.com> * coordinator: add the operator cost time in log field (#1748) Signed-off-by: nolouch <nolouch@gmail.com>
@nolouch I think no. |
STATUS: It seems to be better merged after #1668. You can review it now and give comments.
What problem does this PR solve?
Currently, the safety of timestamp allocation relies on the leader flag be updated in time when lease timeout.
There is a small possibility (for instance, process paused, runtime schedule slow) that the watch channel does not notify in time, which may cause 2 PD servers serve timestamp at the same time. Then the transaction may become corrupted.
Another issue is when a PD server becomes leader, it won't serve timestamps until the leader flag is set. But the leader flag is set after loading all regions from storage, which may take considerably long time for a large cluster.
This PR fixes #1661 and part of #1658
What is changed and how it works?
Introduce
LeaderLease
to periodically renew lease and update expire time.Timestamp service only relies on the lease expire time.
Note that if 2
time.Time
s both have monotonic part, comparison of them will be evaluated by monotonic time, which is not affected by wall clock changes.Check List
Tests
Side effects
Related changes