Use a custom tesla middleware instead of adapter helper function +
custom redirect middleware.
This will also fix "Client died before releasing the connection"
messages when the request pool is overloaded. Since the checkout is
now done after passing ConcurrentLimiter.
This is technically less efficient, since the connection needs to be
checked in/out every time the middleware is left or entered respectively.
But I don't think the nanoseconds we might lose on redirects
to the same host are worth the complexity.
- fix for gun worker termination in some circumstances
- pool for http clients (ex_aws, tzdata)
- default pool timeouts for gun
- gun retries on gun_down messages
- s3 upload timeout if streaming enabled
This patch refactors gun pooling to use Elixir process registry and
simplifies adapter option insertion.
Having the pool use process registry instead of a GenServer has a number of advantages:
- Simpler code: the initial implementation adds about half the lines of code it deletes
- Concurrency: unlike a GenServer, ETS-based registry can handle multiple checkout/checkin
requests at the same time
- Precise and easy idle connection clousure: current proposal for closing idle connections in
the GenServer-based pool needs to filter through all connections once a minute and compare their
last active time with closing time. With Elixir process registry this can be done
by just using `Process.send_after`/`Process.cancel_timer` in the worker process.
- Lower memory footprint: In my tests `gun-memory-leak` branch uses about 290mb on peak load (250 connections)
and 235mb on idle (5-10 connections). Registry-based pool uses 210mb on idle and 240mb on peak load
Haelwenn (lanodan) Monnier