concurrency: unreplicated locks don't provide isolation through commit time

As described in issue #111536, locks themselves don't provide isolation up to the point of commit time because once committed the lock is completely gone. Thus any concurrently committing transaction at a lower timestamp will not observe that the key was locked at the timestamp.

For replicated locks we solved this by always updating the timestamp cache when releasing a lock. The timestamp cache then ensures that any concurrently writing transaction cannot write at a time lower than the commit timestamp when we held a lock.

We don't do the same for unreplicated locks. Thus, even an unreplicated lock is held all the way until commit time, it may not provide isolation up to the commit timestamp.

While we could also update the timestamp cache for all unreplicated locks, this has a few items that we should consider:

Unreplicated locks are not currently reliable anyway since they can be lost. Thus, we would be doing more work for every unreplicated lock and still would not be able to provide an isolation guarantee. See concurrency: unreplicated lock loss detection #142977 for further discussion.

Without further work, this would also require disabling 1PC for any transaction that takes out a unreplicated lock.

cockroach/pkg/kv/kvclient/kvcoord/txn_interceptor_committer.go

Lines 555 to 595 in d6ca127

    
           // Replicated locks must be held until and provide protection up till their 
        
           // transaction's commit timestamp[1]. We ensure this by bumping the timestamp 
        
           // cache to the transaction's commit timestamp for all locked keys when 
        
           // resolving locks. Let's consider external and local replicated locks 
        
           // separately: 
        
           // 
        
           // - External locks: 1PC transactions do not write a transaction record. This 
        
           // means if any of its external locks are resolved by another transaction 
        
           // they'll be resolved as if the transaction were aborted, thus not providing us 
        
           // protection until the transaction's commit timestamp. 
        
           // - Local locks: we have all the information to locally resolve replicated 
        
           // locks and bump the timestamp cache correctly if we're only dealing with local 
        
           // replicated locks. However, the mechanics of 1PC transactions prevent us from 
        
           // hitting it in the common case, where we're acquiring a replicated lock and 
        
           // writing to the same key. 1PC transactions work by stripping the batch of its 
        
           // EndTxnRequest and running it as a non-transactional batch. This means that 
        
           // without some elbow grease, 1PC is bound to fail when it discovers its own 
        
           // replicated lock. For now, we disable 1PC on the client for local locks as 
        
           // well -- this can be optimized in the future. 
        
           // TODO(arul): file an issue about improving things for local locks. 
        
           // 
        
           // [1] This distinction is currently moot for serializable transactions, as they 
        
           // refresh all their reads (locked and unlocked) before committing. Doing so 
        
           // bumps the timestamp cache. However, one can imagine a world where 
        
           // serializable transactions do not need to refresh keys they acquired 
        
           // replicated locks on. In such a world, we would be relying on lock resolution 
        
           // to bump the timestamp cache to the commit timestamp of the transaction. 
        
           func (tc *txnCommitter) maybeDisable1PC(ba *kvpb.BatchRequest) { 
        
           	if tc.disable1PC { 
        
           		return // already disabled; early return 
        
           	} 
        
           	for _, req := range ba.Requests { 
        
           		if readOnlyReq, ok := req.GetInner().(kvpb.LockingReadRequest); ok { 
        
           			_, dur := readOnlyReq.KeyLocking() 
        
           			if dur == lock.Replicated { 
        
           				tc.disable1PC = true 
        
           				return 
        
           			} 
        
           		} 
        
           	} 
        
           }

Jira issue: CRDB-48612

Epic CRDB-49088

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

	// Replicated locks must be held until and provide protection up till their
	// transaction's commit timestamp[1]. We ensure this by bumping the timestamp
	// cache to the transaction's commit timestamp for all locked keys when
	// resolving locks. Let's consider external and local replicated locks
	// separately:
	//
	// - External locks: 1PC transactions do not write a transaction record. This
	// means if any of its external locks are resolved by another transaction
	// they'll be resolved as if the transaction were aborted, thus not providing us
	// protection until the transaction's commit timestamp.
	// - Local locks: we have all the information to locally resolve replicated
	// locks and bump the timestamp cache correctly if we're only dealing with local
	// replicated locks. However, the mechanics of 1PC transactions prevent us from
	// hitting it in the common case, where we're acquiring a replicated lock and
	// writing to the same key. 1PC transactions work by stripping the batch of its
	// EndTxnRequest and running it as a non-transactional batch. This means that
	// without some elbow grease, 1PC is bound to fail when it discovers its own
	// replicated lock. For now, we disable 1PC on the client for local locks as
	// well -- this can be optimized in the future.
	// TODO(arul): file an issue about improving things for local locks.
	//
	// [1] This distinction is currently moot for serializable transactions, as they
	// refresh all their reads (locked and unlocked) before committing. Doing so
	// bumps the timestamp cache. However, one can imagine a world where
	// serializable transactions do not need to refresh keys they acquired
	// replicated locks on. In such a world, we would be relying on lock resolution
	// to bump the timestamp cache to the commit timestamp of the transaction.
	func (tc txnCommitter) maybeDisable1PC(ba kvpb.BatchRequest) {
	if tc.disable1PC {
	return // already disabled; early return
	}
	for _, req := range ba.Requests {
	if readOnlyReq, ok := req.GetInner().(kvpb.LockingReadRequest); ok {
	_, dur := readOnlyReq.KeyLocking()
	if dur == lock.Replicated {
	tc.disable1PC = true
	return
	}
	}
	}
	}

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions