Nginx was closing the client connection after exactly 60 seconds on requests
like clip_video (ffmpeg, 5-40 min) because send_timeout and proxy_send_timeout
both default to 60s. proxy_read_timeout was already 2100s but the other two
timeouts were still at defaults.
With proxy_buffering off, Nginx holds the write path to the client open as soon
as the upstream connection is established. If the upstream sends no data for 60s
(e.g. ffmpeg processing), Nginx treats the idle write path as stalled and closes
the client connection, logging 499 (Client Closed Request).
Fixed: raise proxy_send_timeout and send_timeout to 2100s to match
proxy_read_timeout in the main location block.
Also raised the Gunicorn default timeout from 30s to 120s in gunicorn_conf.py
as a belt-and-suspenders measure (AE_API_GUNICORN_TIMEOUT env var takes precedence).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Newer gunicorn (post-23.0.0) added _get_control_socket_path() which
calls os.path.isabs() on the value — crashing when it is None.
Point the socket to /dev/shm (already used for worker_tmp_dir) so it
is writable inside the container and satisfies the new gunicorn code.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Gunicorn 25.1.0+ enables a 'control_socket' by default, which creates
a root-owned 'gunicorn.ctl' file in the chdir directory. When this
directory is a volume mount (as in our dev/test setups), it causes
permission errors during Docker build context gathering.
This change explicitly sets 'control_socket = None' to prevent the
creation of this file.
