Support load-based replica read #675
Conversation
sticnarf
force-pushed
the
load-based-replica-read
branch
10 times, most recently
from
fb2f1cf to
6e2bfd8
Compare
sticnarf
force-pushed
the
load-based-replica-read
branch
from
6e2bfd8 to
d3c5cfc
Compare
Signed-off-by: Yilin Chen <sticnarf@gmail.com>
sticnarf
force-pushed
the
load-based-replica-read
branch
from
d3c5cfc to
38d5bc8
Compare
txnkv/txnsnapshot/scan.go
Outdated
| @@ -233,6 +233,7 @@ func (s *Scanner) getData(bo *retry.Backoffer) error { | |||
| ResourceGroupTag: s.snapshot.mu.resourceGroupTag, | |||
| RequestSource: s.snapshot.GetRequestSource(), | |||
| ResourceGroupName: s.snapshot.mu.resourceGroupName, | |||
| BusyThresholdMs: uint32(s.snapshot.mu.busyThreshold.Milliseconds()), | |||
There was a problem hiding this comment.
I don't understand why it is not protected by the RWLock?
There was a problem hiding this comment.
Sorry, I think the code before was actually incorrect. Context should not be set here. It should be set below in L243 instead.
Signed-off-by: Yilin Chen <sticnarf@gmail.com>
internal/locate/region_request.go
Outdated
| // exceeding maxReplicaAttempt | ||
| // +-------------------+ || RPC failure && unreachable && no forwarding | ||
| // exceeding maxReplicaAttempt | ||
| // +-------------------+ || RPC failure && unreachable && no forwarding | ||
| // |
There was a problem hiding this comment.
I don't find how to let go fmt ignore it. I add a bar at the line beginning to work around it.
There was a problem hiding this comment.
You can leave an empty line after the graph and before the following code, so that the formatter won't break the indents.
Signed-off-by: Yilin Chen <sticnarf@gmail.com>
Signed-off-by: Yilin Chen <sticnarf@gmail.com>
Signed-off-by: Yilin Chen <sticnarf@gmail.com>
| if ctx != nil && ctx.Store != nil { | ||
| ctx.Store.markAlreadySlow() | ||
| if serverIsBusy := regionErr.GetServerIsBusy(); serverIsBusy != nil { | ||
| if s.replicaSelector != nil { |
There was a problem hiding this comment.
If the replicaSelector is nil would the markAlreadySlow be skipped unexpectedly?
There was a problem hiding this comment.
Actually, this is impossible for TiKV requests. RegionRequestSender always creates a new replicaSelector if not exists.
|
A question about |
Yes, it's a future plan. It's not implemented it in TiKV now because I haven't found a good way to get it efficiently. The gRPC threads don't have access to the engine Maybe it's possible to move local reader to gRPC threads, but there do have a lot to consider. |
Refers to tikv/rfcs#105.
Corresponding TiDB PR: pingcap/tidb#40742