Reduce GIL usage - #376
Open
fpacifici wants to merge 4 commits into
Open
Reduce GIL usage #376fpacifici wants to merge 4 commits into
fpacifici wants to merge 4 commits into
Conversation
Five steps destructured the payload with `let ... else` or `if let`, so a new enum variant would compile without error and silently take the else branch: the header filter would stop filtering, map would become a no-op, routing would be skipped, watermarks would stop tracking the last message time, and the Python filter would panic on the first message. Spell out every arm instead, so adding a variant is a build failure. No behaviour change. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`RawMessage.payload` becomes `Arc<[u8]>` so cloning a message stays cheap: `Broadcaster` clones once per downstream branch, and an upcoming Rust-native payload variant will be cloned along the whole chain. It also makes the immutability this module already documents real. Alongside it, remove code that has no callers: - `#[pyo3(set)]` on `RawMessage.timestamp`/`.schema`. Nothing assigns them from Rust or Python, and rust_streams.pyi already declares both read-only. - the unused `StreamingMessage` enum. - `unwrap_payload()` is narrowed to `#[cfg(test)]`; production code matches on every variant instead. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`RoutedValuePayload::RustRawMessage` carries a `RawMessage` that lives in Rust memory, so a step that can do its work natively never takes the Gil. Every match on the enum gets an arm for it: - watermark and the header filter read the timestamp and the headers natively. Both run before any filtering, on 100% of the traffic; the header filter also stops cloning the header vector per message. - batch stores the variant as it arrives (`Vec<RoutedValuePayload>`, mixed) and only builds the Python list on flush. - sinks and the GCS writer read the bytes natively instead of copying them out of a PyBytes. - map and PythonAdapter convert and ratchet: they already replace the payload, so the Python form is what continues downstream. - filter and router convert transiently and forward the original Rust form, so a following Rust step keeps the cheap path. Nothing emits the variant yet; flipping the source is the next commit. Test builders default to the Rust representation, so the existing suite exercises what production will run, with an explicit Python-representation test for each step that branches on it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`to_routed_value` no longer allocates a `Py<RawMessage>` per message. The payload stays in Rust memory until a step hands it to Python code. On a pipeline shaped like header_filter -> batch -> Python work, where the header filter drops most events, the source, the watermark step and the filter now run without taking the Gil at all. A message the filter drops never enters Python memory, so it costs zero Gil acquisitions instead of three. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
![sentry[bot]](https://avatars.githubusercontent.com/in/12637?s=60&v=4){kind=small}
Comment on lines
+524
to
+526
| fn from(value: &RawMessage) -> Self { | ||
| PyStreamingMessage::RawMessage { | ||
| content: traced_with_gil!(|py| into_pyraw(py, value.clone()).unwrap()), |
There was a problem hiding this comment.
Bug: The use of .unwrap() when converting a RawMessage to a Python object can cause a panic if the object allocation fails, for example due to an out-of-memory error.
Severity: MEDIUM
Suggested Fix
Replace the .unwrap() calls with proper error handling. The From implementations should propagate the PyResult from into_pyraw instead of panicking. This will allow the calling context to handle potential allocation failures gracefully, for instance by returning an error that can be managed further up the call stack.
Prompt for AI Agent
Review the code at the location below. A potential bug has been identified by an AI
agent. Verify if this is a real issue. If it is, propose a fix; if not, explain why it's
not valid.
Location: sentry_streams/src/messages.rs#L524-L526
Potential issue: The `From<&RawMessage>` implementations for `Py<PyAny>` and
`PyStreamingMessage` use `.unwrap()` on the result of the `into_pyraw` function. This
function can fail during Python object creation, for example, in an out-of-memory (OOM)
scenario. Because the conversion from `RawMessage` to a Python object is a common
operation in production pipelines (e.g., in map, filter, and router steps), a failure in
`Py::new()` will result in an unhandled panic, causing the process to crash instead of
handling the error gracefully.
Did we get this right? 👍 / 👎 to inform future reviews.
untitaker
approved these changes
Sep 14, 2026
This branch has not been deployed
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.
The rust consumer moves the payload of each message into python memory immediately after getting the message from Kafka.
This was ok when all messages are processed by Python. Taking the GIL and moving memory was irrelevant.
Now we do a number of processing steps directly in rust like header filters and batching.
Though we did not remove the python processing that happens before filtering.
In prod we observed consumers underperforming when they filter out most messages.
Taking the GIL for each message that is filtered out could be the reason.
This PR introduces Raw rust messages processable by rust native steps without moving data to
Python first.