This is a test repository that allows me to compare the official Elasticsearch client and a minimalist asyncio client.
The idea is to map out code structure and identify possible bottlenecks by profiling some REST API operations (starting with GET /_nodes/_all), and figure out how an asyncio/aiohttp implementation might differ.
The aioelastic folder contains a shim client based on my aioazstorage wrappers that does the bare minimum required to talk to the server but tries to a) parse replies as fast as possible (using ujson) and provide a fully async interface (through an async generator for enumerating nodes).
Both folders (official-client and aioelastic) have a Makefile to provision (make deps), test (make test) and generate PNG profiles (make profile) after running 10K API calls.
These were generated on an i5 iMac with Python 3.7.4 (installed via brew and pyenv) against the official Elasticsearch Docker container, running a single node.
Besides using feedparser (which is interesting) the official client spends most of its time inside recv_into:
aioelastic does some date formatting and uuid generation upon every request (which I find invaluable for tracing requests), but otherwise spends most of its time inside select.kqueue (since I'm using uvloop):
