[client] Use single httpx client for requests & parallelize them

[jstriebel]

Currently, the generated client uses a fresh httpx client for every request (see also issue https://github.com/openapi-generators/openapi-python-client/issues/202 about this), and the resumable-client currently also for one session (but accepts a client in the constructor). This client should also add a global response event hook (see the raise_on_4xx_5xx example) for failed requests. Additionally, the api should use parallel requests where possible e.g. via async-routines.

[motybz]

I vote for this feature since it will make a significant performance improvement on multiple calls. Any plan to implement it?

[jstriebel]

@motybz Sorry for the late reply, this is definitely on our radar! Atm we're focusing on a stable API, but we'll tackle this afterwards to improve up- & download speed.

[Olian04]

@jstriebel what's the status on this issue? Any updates in the past nine months?

[jstriebel]

@Olian04 So far there was little progress, as this is mostly blocked by https://github.com/openapi-generators/openapi-python-client/issues/202.

The following points can be tackled independently already, we'll try to tackle those soonish:

• [x] Register a global httpx Client
• [x] Use this client where httpx is used directly:
  • dataset upload
  • task creation
  • annotation upload
• ~[ ] Add useful middle ware, such as good error handling and token-reloading~ We are using raise_for_status now
• [ ] Use async methods to schedule requests in parallel where possible (mostly dataset download)
• [ ] Test if fsspec can use an httpx client instead of aiohttp (for zarr-streaming)

[dbanty]

Hey @jstriebel, I'm just going over mentions of the openapi-python-client issue to find folks to review https://github.com/openapi-generators/openapi-python-client/pull/775 before it merges (since it's a breaking change). I'd love your input!

[philippotto]

A short update: @fm3 pushed 00e168323a0ac7d47d17edf209ac0d8a340f3b34 last week which removes the auto-generated client. As a byproduct, allocating a fresh httpx client for each api call is not done, anymore. Parallelizing the download of a dataset is still an open point.

[normanrz]

The webknossos package now uses zarrita under the hood which uses async io to fetch chunks concurrently.