@dadjeibaah Thanks for this change, happy to see it in linkerd!

@adleong absolutely! I'm back on Tuesday however, if you think it is worth waiting till then I'd be happy to look at the change and give it a try. But I believe the only thing to discuss here is the default value of the parameter (I don't really like finagle's default 5 MiB), and it is not that important anyway.

Btw we just started using similar patch (based on linkerd 1.5.2) at NCBI.

According to my benchmarks the change also improves latency for fixed length requests. On my scenarios the difference is about 10-15% for requests below 1 MiB but it gets much more pronounced as request size grows larger or link to linkerd gets slower (I think everyone has linkerds on the same host with their apps, so it's all about copying data in memory anyway, hence only request size is a significant factor).

I have only single result on me right now, but if there's interest I could post more when I'm back

baseline
Client.invoke:invoke·p0.95               4096  sample           5.677           ms/op

vanilla l5d
Client.invoke:invoke·p0.95               4096  sample          10.879          ms/op

patched l5d
Client.invoke:invoke·p0.95               4096  sample           8.282          ms/op

Where baseline is single threaded direct invocation of a service on localhost which just counts bytes in request. Linkerd is executed on Java 8 with -Xms256m -Xmx256m. This case is 4MiB request.