Update trace summary script

[pskiran1]

No description provided.

[Tabrizian]

I feel like still we should be looking for ways to correctly record GRPC WAITREAD or remove this value.

One scenario that I can imagine being confusing for the user is when they compare the value of GRPC infer request in gRPC and HTTP and notice much longer values for gRPC since it takes into account the gaps between the requests as well.

[tanmayv25]

@pskiran1 Were you able to investigate whether gRPC server allows us to record actual request read time? Similar to libevhtp which allow us to record a timestamp on the first set of bytes, does gRPC have any hooks which can be used to detect the network activity - hence the beginning of a read from the network?

If not, then I would advocate for removing the value as well to avoid any confusion.

[pskiran1]

As we discussed internally, gRPC does not provide a direct hook for recording request read time, and using alternative methods such as interceptors may add latency/overhead to the current server. Therefore, we have decided to remove the GRPC WAITREAD related values in the summary to avoid confusion.

[tanmayv25]

a small change otherwise looks good to me.

[Tabrizian]

@pskiran1 Actually, I realized that you need to remove it from grpc trace recording too.

[Tabrizian]

server/src/grpc/infer_handler.cc

Line 667 in 8fc7b10

std::make_pair("GRPC_WAITREAD_START", TraceManager::CaptureTimestamp()));