Add Python SDK option to opt out of HTTP/2

[sarmientoF]

Summary

• Keep HTTP/2 as the Python SDK default.
• Add a public http2 connection option so callers can opt out with http2=False for API and envd transports.
• Thread the option through sync/async API clients and sandbox envd transport creation.
• Add tests for the default HTTP/2 behavior and the explicit HTTP/1.1 opt-out.

Why

For normal use, HTTP/2 remains the right default. The opt-out is for server-side high fan-out workloads where httpx uses one HTTP/2 connection per origin and therefore one server MAX_CONCURRENT_STREAMS budget. Once a workload has more long-lived concurrent envd RPCs than the peer stream cap, HTTP/2 requests queue behind that single connection instead of spreading across the existing connection pool.

http2=False gives those workloads a documented alternative: HTTP/1.1 uses the SDK/httpx connection pool and avoids coupling all concurrent envd RPCs to one HTTP/2 stream budget. This is useful for bash/filesystem/commands/pty/control-plane fan-out and does not change behavior for default users.

Reproduction / metrics

Using this branch and AsyncSandbox.commands.run, I ran 180 concurrent commands against one sandbox:

python /tmp/e2b_cloud_http2_repro.py \
  --env-file /Users/fernando-sarmiento/Desktop/jai/E2B/.env \
  --concurrency 180 \
  --sleep-seconds 4 \
  --request-timeout 45

Workload: sleep 4; echo ok, 180 concurrent server-streaming envd RPCs.

mode	ok/errors	elapsed	throughput	p50	p95	p99	max
HTTP/2 default	180/0	9.03s	19.94/s	4.82s	9.01s	9.01s	9.01s
HTTP/1.1 opt-out	180/0	5.20s	34.63/s	5.01s	5.16s	5.17s	5.18s

The HTTP/2 run completes in two waves because concurrency above the stream budget cannot use a second HTTP/2 connection. The HTTP/1.1 opt-out completes in one wave because requests are distributed through the connection pool.

Tests

• ./.venv/bin/pytest tests/test_api_client_transport.py tests/sync/sandbox_sync/test_config_propagation.py tests/async/sandbox_async/test_config_propagation.py
• ./.venv/bin/ruff check .
• ./.venv/bin/ruff format --check .
• ./.venv/bin/ty check

[cla-bot]

We require contributors to sign our Contributor License Agreement, and we don't have @sarmientoF on file. You can sign our CLA at https://e2b.dev/docs/cla . Once you've signed, post a comment here that says '@cla-bot check'

[changeset-bot]

🦋 Changeset detected

Latest commit: ca635b4

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
@e2b/python-sdk	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[cla-bot]

We require contributors to sign our Contributor License Agreement, and we don't have @sarmientoF on file. You can sign our CLA at https://e2b.dev/docs/cla . Once you've signed, post a comment here that says '@cla-bot check'

[sarmientoF]

@cla-bot check

[cla-bot]

The cla-bot has been summoned, and re-checked this pull request!

[matthewlouisbrockman]

switching to http1 leads to more problems than it solves

[matthewlouisbrockman]

To reopen, please submit a reproduction of the issue along with metrics for performance

[sarmientoF]

I revised the branch based on the feedback:

• HTTP/2 remains the default.
• The PR now only adds a public http2=False opt-out and threads it to API/envd transports.
• I added the requested reproduction/metrics to the PR body.

Key result from a real E2B sandbox workload: 180 concurrent AsyncSandbox.commands.run("sleep 4; echo ok") calls completed with HTTP/2 default p95 ~= 9.01s versus HTTP/1.1 opt-out p95 ~= 5.16s. This demonstrates the concrete case for HTTP/1.1: high fan-out, long-lived envd RPCs above the HTTP/2 stream budget serialize on one httpx HTTP/2 connection, while HTTP/1.1 spreads across the connection pool.

This keeps default users on HTTP/2 and gives high fan-out users a documented escape hatch instead of requiring monkeypatching.

[sarmientoF]

Opened a fresh PR with the revised scope and current head SHA because this closed PR would not reopen or refresh its head ref: #1407

The new PR keeps HTTP/2 as default and includes the reproduction/metrics requested here.