fix the issue that health check may set liveness wrongly #1127
Conversation
Signed-off-by: zyguan <zhongyangguan@gmail.com>
Signed-off-by: zyguan <zhongyangguan@gmail.com>
Signed-off-by: zyguan <zhongyangguan@gmail.com>
Signed-off-by: zyguan <zhongyangguan@gmail.com>
Signed-off-by: zyguan <zhongyangguan@gmail.com>
|
@cfzjywxk PTAL |
Signed-off-by: zyguan <zhongyangguan@gmail.com>
Signed-off-by: zyguan <zhongyangguan@gmail.com>
Signed-off-by: zyguan <zhongyangguan@gmail.com>
|
/cc @crazycs520 PTAL |
| newStore := &Store{storeID: s.storeID, addr: addr, peerAddr: store.GetPeerAddress(), saddr: store.GetStatusAddress(), storeType: storeType, labels: store.GetLabels(), state: uint64(resolved)} | ||
| newStore.livenessState = atomic.LoadUint32(&s.livenessState) | ||
| newStore.unreachableSince = s.unreachableSince |
There was a problem hiding this comment.
How about abstracting a newStore function to force specify the livenessState and unreachableSince as the input parameters.
There was a problem hiding this comment.
How about to handle it later (as a part of #1104). There are 4 &Store{...} in region_cache.go now, I'd like to handle them togather according to the store lifecycle. Let this PR fix the corresponding issue by now.
| @@ -2783,50 +2785,85 @@ func (s *Store) requestLivenessAndStartHealthCheckLoopIfNeeded(bo *retry.Backoff | |||
| // It may be already started by another thread. | |||
| if atomic.CompareAndSwapUint32(&s.livenessState, uint32(reachable), uint32(liveness)) { | |||
| s.unreachableSince = time.Now() | |||
| go s.checkUntilHealth(c) | |||
| reResolveInterval := 30 * time.Second | |||
There was a problem hiding this comment.
Better to make it a constant value.
There was a problem hiding this comment.
The rawkv tests use failpoint to set it shorter.
There was a problem hiding this comment.
how about
var DefReResolveInterval = 30 * time.Second // global scope.
func (s *Store) checkUntilHealth(c *RegionCache, liveness livenessState){
reResolveInterval = DefReResolveInterval
...
}and if test needs a shorter reResolveInterval, change DefReResolveInterval directly, no need to use the following failpoint.
There was a problem hiding this comment.
how about
var DefReResolveInterval = 30 * time.Second // global scope. func (s *Store) checkUntilHealth(c *RegionCache, liveness livenessState){ reResolveInterval = DefReResolveInterval ... }and if test needs a shorter reResolveInterval, change
DefReResolveIntervaldirectly, no need to use the following failpoint.
Personally I do not prefer that. The global var is required to be handled carefully to avoid data race in ut. I've struggled with SetRegionCacheTTLSec in #1122 for about an hour.
| logutil.BgLogger().Info("[health check] store meta deleted, stop checking", zap.Uint64("storeID", s.storeID), zap.String("addr", s.addr)) | ||
| if s.getResolveState() == deleted { | ||
| // if the store is deleted, a new store with same id must be inserted (guaranteed by reResolve). | ||
| newStore, _ := c.getStore(s.storeID) |
There was a problem hiding this comment.
Is it necessary to add an assertion newStore != nil here?
There was a problem hiding this comment.
It's guaranteed by reResolve, changeToActiveStore also rely on it.
Signed-off-by: zyguan <zhongyangguan@gmail.com>
* fix the issue that health check may set liveness wrongly Signed-off-by: zyguan <zhongyangguan@gmail.com> * fix lint issue Signed-off-by: zyguan <zhongyangguan@gmail.com> * fix rawkv ut Signed-off-by: zyguan <zhongyangguan@gmail.com> * fix data race Signed-off-by: zyguan <zhongyangguan@gmail.com> * use getStore instead of accessing storeMu directly Signed-off-by: zyguan <zhongyangguan@gmail.com> * make TestAccessFollowerAfter1TiKVDown stable Signed-off-by: zyguan <zhongyangguan@gmail.com> * make TestBackoffErrorType stable Signed-off-by: zyguan <zhongyangguan@gmail.com> * address comments Signed-off-by: zyguan <zhongyangguan@gmail.com> --------- Signed-off-by: zyguan <zhongyangguan@gmail.com> Co-authored-by: disksing <i@disksing.com> Signed-off-by: zyguan <zhongyangguan@gmail.com>
* fix the issue that health check may set liveness wrongly * fix lint issue * fix rawkv ut * fix data race * use getStore instead of accessing storeMu directly * make TestAccessFollowerAfter1TiKVDown stable * make TestBackoffErrorType stable * address comments --------- Signed-off-by: zyguan <zhongyangguan@gmail.com> Co-authored-by: disksing <i@disksing.com>
close #1111