fix(inference-gateway): decode compressed middleware responses - #1218
Merged
Conversation
Signed-off-by: Brian Newsom <brnewsom@nvidia.com>
Contributor
📝 WalkthroughWalkthroughThe inference gateway now enables aiohttp upstream content decoding. Decoded content feeds SSE and JSON processing. Payload-decoding and JSON-processing failures return distinct 502 errors. Tests cover headers, gzip JSON, gzip SSE, and failure paths. ChangesUpstream response decoding
Suggested reviewers: Sequence Diagram(s)sequenceDiagram
participant Upstream
participant fetch_proxy_response
participant SSEParser
participant JSONParser
Upstream->>fetch_proxy_response: Encoded response
fetch_proxy_response->>SSEParser: Decoded streaming chunks
fetch_proxy_response->>JSONParser: Decoded buffered body
JSONParser-->>fetch_proxy_response: Parsed object or 502 error
🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Actionable comments posted: 3
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@services/core/inference-gateway/src/nmp/core/inference_gateway/api/proxy.py`:
- Around line 486-499: Update the decompression flow around the chunk loop and
final decoder.flush() to track whether the final compressed member completed
before aiohttp can reset decoder state. Require gzip/deflate
ZLibDecompressor.eof and equivalent end-of-stream completion indicators for
Brotli and Zstandard; raise UpstreamContentDecodingError(content_encoding) for
incomplete bodies, while preserving normal decoded and trailing output.
- Around line 519-524: Update _parse_sse_stream and the surrounding
stream_response_result flow so UpstreamContentDecodingError raised lazily by
_decode_content_chunks or _parse_sse_chunks is handled before SSE output begins,
returning the established 502 response instead of propagating after headers are
sent. Add a regression test covering corrupt gzip content in an SSE response and
verify no partial successful response is emitted.
- Around line 458-471: Update _build_content_decoder to parse comma-separated
Content-Encoding values, ignore surrounding whitespace, and construct a decoder
chain in reverse application order so values such as “gzip, br” decode
correctly. Preserve identity/empty handling and unsupported-coding errors, then
add buffered JSON and fragmented SSE coverage for chained encodings.
🪄 Autofix
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Enterprise
Run ID: a10fde9a-7255-4d13-8dad-a0618ccea8ae
📒 Files selected for processing (4)
services/core/inference-gateway/src/nmp/core/inference_gateway/api/proxy.pyservices/core/inference-gateway/src/nmp/core/inference_gateway/service.pyservices/core/inference-gateway/src/nmp/core/inference_gateway/testing/_loopback.pyservices/core/inference-gateway/tests/unit/test_proxy.py
Contributor
|
Signed-off-by: Brian Newsom <brnewsom@nvidia.com>
Signed-off-by: Brian Newsom <brnewsom@nvidia.com>
Contributor
There was a problem hiding this comment.
Actionable comments posted: 1
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@services/core/inference-gateway/tests/unit/test_proxy.py`:
- Around line 1514-1532: Move the header-sanitization assertions from
test_fetch_proxy_response_decodes_chained_json to a stream_response_result test
that exercises re-serialization. Assert that content-length, content-encoding,
and transfer-encoding are absent from the serialized response headers, while
keeping this fetch_proxy_response test focused on chained JSON decoding and
preserving the direct proxy path’s content-encoding behavior.
🪄 Autofix
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Enterprise
Run ID: bf38e0cc-8b97-430a-83c8-f09c0394241e
📒 Files selected for processing (2)
services/core/inference-gateway/src/nmp/core/inference_gateway/api/proxy.pyservices/core/inference-gateway/tests/unit/test_proxy.py
🚧 Files skipped from review as they are similar to previous changes (1)
- services/core/inference-gateway/src/nmp/core/inference_gateway/api/proxy.py
Signed-off-by: Brian Newsom <brnewsom@nvidia.com>
Signed-off-by: Brian Newsom <brnewsom@nvidia.com>
Contributor
There was a problem hiding this comment.
Actionable comments posted: 2
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@services/core/inference-gateway/src/nmp/core/inference_gateway/api/proxy.py`:
- Line 492: Restore chained Content-Encoding support in the proxy response
handling around auto_decompress=True by explicitly decoding ordered coding
chains or preserving a decoder chain for multi-coding responses, including
combinations such as gzip, deflate. In
services/core/inference-gateway/src/nmp/core/inference_gateway/api/proxy.py
lines 492-492, update the implementation; in
services/core/inference-gateway/tests/unit/test_proxy.py lines 1498-1518, add
coverage for chained JSON responses and fragmented chained SSE responses. Use uv
for Python tooling.
- Around line 519-522: Update the compressed SSE handling in _parse_sse_stream’s
caller so compressed responses are fully buffered and validated before
constructing or returning ResponseResult, while uncompressed SSE remains
live-streamed. Add corrupt compressed-SSE coverage in
services/core/inference-gateway/tests/unit/test_proxy.py:1553-1581 asserting the
gateway returns 502 before any response iterator is exposed; the proxy
implementation change is required at
services/core/inference-gateway/src/nmp/core/inference_gateway/api/proxy.py:519-522.
🪄 Autofix
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Enterprise
Run ID: 9ba3af43-8e0c-46e7-9afc-894c63a1481f
📒 Files selected for processing (4)
services/core/inference-gateway/src/nmp/core/inference_gateway/api/proxy.pyservices/core/inference-gateway/src/nmp/core/inference_gateway/service.pyservices/core/inference-gateway/src/nmp/core/inference_gateway/testing/_loopback.pyservices/core/inference-gateway/tests/unit/test_proxy.py
🚧 Files skipped from review as they are similar to previous changes (2)
- services/core/inference-gateway/src/nmp/core/inference_gateway/service.py
- services/core/inference-gateway/src/nmp/core/inference_gateway/testing/_loopback.py
JashG
approved these changes
Aug 13, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Inference middleware uses the gateway's non-decompressing HTTP session but then parses upstream bytes as JSON or SSE. When a sandbox egress proxy negotiates gzip, that path sees compressed bytes and returns a misleading JSON 502 or an empty event stream. Enable aiohttp decompression only for middleware-aware requests while leaving direct proxy responses byte-transparent.
Changes
auto_decompress=Trueon the middleware-aware upstream request; keep the shared session default disabled for direct proxying.Scope boundary
Content-Encodinghandling to aiohttp instead of maintaining gateway-specific codecs.Accept-Encodingheaders or sandbox proxy configuration; the gateway now handles compression negotiated anywhere on the upstream path.Type of Change
Quality Gates
Verification
Signed-off-by:traileruv run pre-commit run -apasses, or any blocked checks are identified belowTargeted validation:
uv run --frozen pytest services/core/inference-gateway/tests/unit/test_proxy.py -k 'fetch_proxy_response_decodes_gzip' -vfailed both regression cases with the reported JSON 502 and an empty SSE event list.uv run --frozen pytest services/core/inference-gateway/tests/unit -v— 463 passed, 5 skipped.uv run --frozen pre-commit run -a— all hooks passed.make refresh-openapi— passed with no generated changes.external-onlysbx sandbox against this branch's local gateway.HTTP_PROXY/HTTPS_PROXYwere set,host.docker.internalwas absent from bothNO_PROXYandno_proxy, and the run command did not override either bypass list. Gateway chat-completion requests returned 200 rather than the reported non-JSON 502.run.json.status=completedwithwinner_agent=agent-1and validationreward=1.0,shape_ok=1.0. One unrelated upstream request stalled with the gateway's existing unbounded timeout; interrupting and rerunning the documented command with the same experiment directory resumed successfully from persisted state.uv run --frozen pytest plugins/nemo-experimentalist/tests/experimentalist/ -k smoke -q— 76 passed, 15 skipped, 321 deselected.