I just ran another experiment with the patched ShouldGoInCache function. Unfortunately, results don’t look to promising.
This shows Dgraph client latency within the 5 minutes that I ran the patched binary. v23 is on the left y-axis, the v24 modified version is on the right y-axis. As you can see, where latency for v23 is measured in ms, the latency for the modified v24 is measured in seconds. Avg for v24 is 500ms, avg for v23 is 22ms.
As discussed, this is the output of dgraph debuginfo with the modified version of v24 running in production. It doesn’t include the zero files, nor the dgraph trace file, since it’s bigger then the max file limit that I can upload here (222MB). If you are interested in this file as well, I can try to provide it through different means.
alpha_block.gz (718 Bytes)
alpha_jemalloc.gz (193 Bytes)
alpha_metrics.gz (174.1 KB)
alpha_mutex.gz (719 Bytes)
alpha_state.gz (305.5 KB)
alpha_threadcreate.gz (890 Bytes)
alpha_cpu.gz (90.7 KB)
alpha_goroutine.gz (1.4 MB)
alpha_health.gz (185 Bytes)
alpha_heap.gz (82.2 KB)