[DNM][WIP] client/leasemanager: create a package to build transaction… #45156
Conversation
|
This change is |
There was a problem hiding this comment.
Reviewable status:
complete! 0 of 0 LGTMs obtained (waiting on @ajwerner)
pkg/kv/txn_coord_sender.go, line 882 at r1 (raw file):
defer tc.mu.Unlock() if tc.mu.txn.WriteTimestamp.Less(ts) { tc.mu.txn.WriteTimestamp = ts
does this ensure commit at >= WriteTimestamp even if the txn does no more writes?
There was a problem hiding this comment.
Reviewable status:
pkg/kv/txn_coord_sender.go, line 882 at r1 (raw file):
Previously, sumeerbhola wrote…
does this ensure commit at >=
WriteTimestampeven if the txn does no more writes?
Yes. The timestamp for the EndTxnRequest comes from the WriteTimestamp
There was a problem hiding this comment.
Reviewable status:
pkg/kv/txn_interceptor_heartbeater.go, line 238 at r2 (raw file):
func (h *txnHeartbeater) forceHeartbeat() error { h.forceHeartbeatCh <- struct{}{}
The synchronization here isn't clear. It seems like you can't call this more than once. Make that clear from the API?
I sort of also wish this call and its parents took a context and were cancelable.
I think we can come up with fancier dances with channels to make this reentrant and deal with context cancelation.
sumeerbhola
force-pushed
the
ajwerner/txn-scoped-leases
branch
from
64956b1
to
3676f87
Compare
… scoped leases This is a very WIP implementation of transaction scoped locks. The interface is garbage and not very considered. Nevertheless it seems to mostly work. The biggest problem I see so far is that at high concurrency there are cases where a lock can lay down all of the relevant intents and then still fail to commit. This happens because the TS cache information is lost due to a lease transfer before the txn record is layed down. Such a possibility is a big problem for the lease use cases which assume that an intent will be valid until the transaction expires. To some extent this problem makes sense given the transaction expiration isn't well defined before the txn record has been written. Release note: None
- Txn.ForceHeartbeat() writes a transaction record. - Txn.ExpiryTimestamp() returns the current expiry timestamp. - TODOs for plumbing a special Intent that updates the ts cache when removed due to abort. Release note: None
More plumbing for special intent that updates ts cache. Remaining work is to somehow update the ts cache in cmd_resolve_intent.go and cmd_resolve_intent_range.go, Release note: None
Plumbing to update the ts cache when removing a "leasing intent". Release note: None
yuzefovich
force-pushed
the
ajwerner/txn-scoped-leases
branch
from
3676f87
to
e8d6ae1
Compare
Added proof sketch for lease/lock implementation. Added description of alternative that does not have contention between concurrent acquisitions of shared lease. Release note: None
… scoped leases
This is a very WIP implementation of transaction scoped locks.
The interface is garbage and not very considered.
Nevertheless it seems to mostly work. The biggest problem I see so far is that
at high concurrency there are cases where a lock can lay down all of the
relevant intents and then still fail to commit. This happens because the TS
cache information is lost due to a lease transfer before the txn record is layed
down. Such a possibility is a big problem for the lease use cases which assume
that an intent will be valid until the transaction expires. To some extent this
problem makes sense given the transaction expiration isn't well defined before
the txn record has been written.
Release note: None