Opt-in core-to-REST-server instrumentation
Provides crucial information on core-to-REST-server HTTP operations. This is essential for analyzing and minimizing remote-request latencies. An indispensable counterpart to Jaeger tracing, while Jaeger tracing isn't enough to give us a full picture on all interactions in all contexts. Having this in-core will make these performance numbers readily accessible where they're most needed, namely, in notebook environments.
TYPE: IMPROVEMENT DESC: Opt-in core-to-REST-server instrumentation
@ypatia is there anything else to do on this PR? If not can we approve it?
@ypatia is there anything else to do on this PR? If not can we approve it?
It's ok for me, but let's have an approval from either @Shelnutt2 or @ihnorton before we merge
As per @Shelnutt2's initial review we have:
- Removed
flush() - Removed the config variable all along as it was only needed to disable
flush()if that would bring perf degradation. Now only trace level logging level is needed. - Replaced
LOG_TRACE()macro w/ using the structuredlogger_member ofCurlclass to give finer debug info on what query is executing.
🚀