Repository navigation
feat(kernel): expose HTTP/2 opt-out - #1070
aakash-saravanan-db wants to merge 9 commits into
Conversation
There was a problem hiding this comment.
🔀 Cross-reviewed: two independent reviews (Claude Opus 5.5 · GPT-6.1 Terra), checked by a judge model (Claude Sonnet 5.5).
Verdict: 1 High · 2 Low
Mostly good, with 1 medium concern. Leaving enable_http2 unset keeps working on older wheels, but setting it explicitly on a wheel that lacks the option fails with a vague error instead of a clear NotSupportedError.
This overview is Claude Opus 5.5's initial take; GPT-6.1 Terra added findings it doesn't mention — see the findings below.
There was a problem hiding this comment.
🔀 Cross-reviewed: two independent reviews (Claude Opus 5.5 · GPT-6.1 Terra), checked by a judge model (Claude Sonnet 5.5).
Verdict: 1 High · 1 Medium
Mostly good, with 1 medium concern: an explicit enable_http2 reaches older kernel wheels without the capability check that geospatial uses. Validation and forwarding look correct. Nit: the new README paragraph sits between the install heading and its install line.
This overview is Claude Opus 5.5's initial take; GPT-6.1 Terra added findings it doesn't mention — see the findings below.
Other findings
- 🟡 Medium — The new tests only capture kwargs in a fake Session, so they cannot detect a native transport that ignores the flag or incorrectly negotiates HTTP/2 through a proxy. Add a real-kernel regression that verifies negotiated protocols for omitted, true and false settings, including proxy fallback. (Found by GPT-6.1 Terra · the judge couldn't confirm it)
There was a problem hiding this comment.
🔀 Cross-reviewed: two independent reviews (Claude Opus 5.5 · GPT-6.1 Terra), checked by a judge model (Claude Sonnet 5.5).
Verdict: 1 High · 1 Medium
The forwarding and validation look correct and are tested, with one medium concern: an explicit enable_http2 is not gated for older kernel wheels.
This overview is Claude Opus 5.5's initial take; GPT-6.1 Terra added findings it doesn't mention — see the findings below.
Other findings
- 🟡 Medium — The new test only captures kwargs in a permissive fake, so CI cannot detect a missing native argument or verify HTTP/2 negotiation, fallback, and proxy opt-out. Add a regression using the matching real kernel that exercises the settings and checks the negotiated transport. (Found by GPT-6.1 Terra · the judge couldn't confirm it)
There was a problem hiding this comment.
🔀 Cross-reviewed: two independent reviews (Claude Opus 5.5 · GPT-6.1 Terra), checked by a judge model (Claude Sonnet 5.5).
Verdict: 1 High · 1 Medium · 1 Low
Mostly fine, with 1 medium concern: an explicit enable_http2 isn't checked against the installed kernel Session signature, so older wheels fail with an unclear error. The plumbing and the bool validation are correct and tested.
This overview is Claude Opus 5.5's initial take; GPT-6.1 Terra added findings it doesn't mention — see the findings below.
Other findings
- 🟡 Medium — The new tests only capture kwargs in a fake Session, so the maintained suite never verifies that the real kernel negotiates HTTP/2 by default or honors the HTTP/1.1 opt-out. Add an in-repo real-kernel regression that observes the negotiated protocols for both settings. (Found by GPT-6.1 Terra · the judge couldn't confirm it)
Signed-off-by: Aakash Saravanan <aakash.saravanan@databricks.com>
Signed-off-by: Aakash Saravanan <aakash.saravanan@databricks.com>
There was a problem hiding this comment.
🔀 Cross-reviewed: two independent reviews (Claude Opus 5.5 · GPT-6.1 Sol), checked by a judge model (Claude Sonnet 5.5).
Verdict: 1 Medium · 1 Low
Mostly good, with 1 medium concern. Forwarding and the bool check are correct, but an explicit enable_http2 isn't checked against what the installed kernel Session accepts, unlike the existing geospatial option.
Signed-off-by: Aakash Saravanan <aakash.saravanan@databricks.com>
There was a problem hiding this comment.
🔀 Cross-reviewed: two independent reviews (Claude Opus 5.5 · GPT-6.1 Sol), checked by a judge model (Claude Sonnet 5.5).
Verdict: 1 Low
Looks good, with 1 low concern. The option is checked for type before session construction, gated through _kernel_session_accepts_kwarg (which fails closed), forwarded only when set, and tested at both the session and client layers.
| # Release History | ||
|
|
||
| # Unreleased | ||
| - Kernel HTTPS traffic negotiates HTTP/2 with HTTP/1.1 fallback. Set `enable_http2=False` to force HTTP/1.1. Configured proxies also force HTTP/1.1. |
There was a problem hiding this comment.
🔵 Low — The CHANGELOG entry says kernel HTTPS negotiates HTTP/2 and that enable_http2=False forces HTTP/1.1. The floor is still ^1.4.0, and the code raises NotSupportedError when the installed wheel lacks the kwarg, so the entry is only true with a newer wheel. Name the required kernel version in the entry, or raise the pyproject floor once that wheel ships.
🔀 Found by Claude Opus 5.5 · confirmed by the judge
There was a problem hiding this comment.
The unsupported-wheel error is now explicit, but the actual kernel requirement and lockfile still need to move to the release providing the option before this PR merges. The changelog will ship with that pin update; this release prerequisite remains open.
There was a problem hiding this comment.
Still open at d517411: CHANGELOG.md:4 doesn't name a kernel version, and the kernel floor and lockfile haven't been raised.
Signed-off-by: Aakash Saravanan <aakash.saravanan@databricks.com>
There was a problem hiding this comment.
🔀 Cross-reviewed: two independent reviews (Claude Opus 5.5 · GPT-6.1 Sol), checked by a judge model (Claude Sonnet 5.5).
Verdict: 1 Low
Looks good: 1 low concern. The HTTP/2 helper correctly rejects non-bool values and checks Session support before construction, and the tests cover omitted, True, False, a non-bool value and an older fixed-signature Session. Nit: the CONNECTION_PARAMETERS.md list of kernel-only options leaves out enable_http2.
| # Release History | ||
|
|
||
| # Unreleased | ||
| - Kernel HTTPS traffic negotiates HTTP/2 with HTTP/1.1 fallback. Set `enable_http2=False` to force HTTP/1.1. Configured proxies also force HTTP/1.1. |
There was a problem hiding this comment.
🔵 Low — The kernel floor is still ^1.4.0, so older wheels may not negotiate HTTP/2 or force HTTP/1.1 behind a proxy as the CHANGELOG and README claim. Explicit enable_http2 already raises NotSupportedError on those wheels. Raise the floor and lockfile to the release that supports enable_http2, or word the notes as depending on that kernel version.
🔀 Found by Claude Opus 5.5 · confirmed by the judge
There was a problem hiding this comment.
Duplicate of #1070 (comment). This PR will merge only with the kernel requirement and lockfile raised to the release providing kernel #483. The installed-wheel capability check handles explicit values in the meantime; the release floor remains a merge prerequisite.
Feature or gap
Expose HTTP/2 negotiation control for kernel connections.
What this PR addresses
Add the boolean
enable_http2connection option. Omission preserves the native default;Falseforces HTTP/1.1. An explicit value requires a supporting installedSessionconstructor and raisesNotSupportedErrorbefore construction when unsupported. Document the option in the connection-parameter table.Dependencies
Requires kernel #483. Before merging, update the kernel requirement and lockfile to the release providing
Session(enable_http2=...); the current^1.4.0floor remains unchanged pending that release.Tests and validation
Focused tests cover omitted/true/false forwarding, invalid types and a fixed-signature older Session. Compatibility and large-result measurements used earlier local builds. Published-artifact and scheduled verification remain release prerequisites.
This pull request and its description were written by Isaac.