StoreAPI batching

[GiedriusS]

Is your proposal related to a problem?

Sending responses one-by-one from StoreAPI is not ideal:

• Compression suffers because compression is per-message
• We have to allocate a lot of small objects hence increased GC pressure

Since each response is small, my proposal is to pack multiple responses into one message.

Describe the solution you'd like

Tunable batch size parameter on the StoreAPI level (request - batch size, response is batched accordingly) and inside of the querier.

Describe alternatives you've considered

N/A

Additional context

N/A

[harry671003]

Is the proposal to batch multiple Series in the Series response?

I'm wondering if GRPC under the hood automatically batches multiple SeriesResponse into one TCP request.

message SeriesResponse {
  oneof result {
    repeated Series series = 1; // Batch?
    string warning = 2;
    google.protobuf.Any hints = 3;
  }
}

thanos/pkg/store/storepb/rpc.proto

Line 175 in b144eba

message SeriesResponse {

[GiedriusS]

It's not about the transport layer but about reducing GC pressure through allocating much fewer storepb.SeriesResponse objects on both sides. Also, it's a prelude to further possible improvements like adding a symbol table. With batch size 1 doing something like that doesn't make sense.