fix(tracing): Fix memory leak in SGP tracing processors

[smoreinis]

Summary

• Fix memory leak in SGPSyncTracingProcessor and SGPAsyncTracingProcessor where self._spans dict grew linearly with every request and never shrank
• on_span_end() used dict.get() (read-only) instead of dict.pop() (read-and-remove), so completed spans were never cleaned up
• shutdown() was the only cleanup path but is never called anywhere in the codebase
• After 100k requests this would accumulate 400k orphaned span objects in memory

Changes

• Changed self._spans.get(span.id) → self._spans.pop(span.id, None) in both SGPSyncTracingProcessor.on_span_end() and SGPAsyncTracingProcessor.on_span_end()
• Added 6 new tests proving the leak existed and is now fixed (3 sync, 3 async)

Test plan

• New tests fail before fix (100 spans leaked after 100 complete lifecycles)
• New tests pass after fix (0 spans after 100 complete lifecycles)
• Full test suite passes (591 passed, 959 skipped)
• Lint clean

Greptile Summary

This PR fixes a genuine memory leak in SGPSyncTracingProcessor and SGPAsyncTracingProcessor by changing self._spans.get(span.id) to self._spans.pop(span.id, None) in each processor's on_span_end() method. The fix is correct, minimal, and well-targeted — completed spans are now removed from self._spans as soon as they finish rather than accumulating indefinitely.

Key changes:

• sgp_tracing_processor.py: .get() → .pop() in both SGPSyncTracingProcessor.on_span_end and SGPAsyncTracingProcessor.on_span_end
• New test file with 6 tests (3 sync / 3 async) covering the full span lifecycle, mid-lifecycle assertion, and the unknown-span no-op path
• Three new __init__.py files to make test directories proper packages

Notes:

• asyncio_mode = "auto" is set in pyproject.toml, so the async test methods run correctly without needing per-test @pytest.mark.asyncio markers
• Both _make_processor helpers now correctly use with patch(...) context managers (not decorator-style), and test bodies re-patch create_span — the pattern is consistent between sync and async classes

Confidence Score: 5/5

Safe to merge — the fix is a minimal, correct one-method change in each processor with no API surface impact.

The core change is a single well-understood substitution (.get() → .pop()) that directly and correctly addresses the described leak. asyncio_mode = auto is confirmed in pyproject.toml so async tests are actually exercised. All remaining observations are pre-existing or already flagged in prior review threads, leaving no new P0/P1 findings in this revision.

No files require special attention.

Important Files Changed

Filename	Overview
src/agentex/lib/core/tracing/processors/sgp_tracing_processor.py	Core bug fix: .get() → .pop() in both on_span_end methods; change is correct and complete. Also removes unnecessary multi-line formatting for warning log strings.
tests/lib/core/tracing/processors/test_sgp_tracing_processor.py	New test file with 6 tests covering sync and async memory-leak scenarios; asyncio_mode = auto ensures async tests run, and patching patterns are correct in both sync and async helpers.

Sequence Diagram

sequenceDiagram
    participant C as Caller
    participant P as SGP[Sync|Async]TracingProcessor
    participant S as self._spans (dict)
    participant SGP as SGPClient

    C->>P: on_span_start(span)
    P->>SGP: create_span(...)
    SGP-->>P: sgp_span
    P->>SGP: flush / upsert_batch
    P->>S: self._spans[span.id] = sgp_span

    Note over S: span lives in dict while active

    C->>P: on_span_end(span)
    P->>S: pop(span.id, None)  ← FIX (was .get())
    S-->>P: sgp_span (removed from dict)
    P->>SGP: flush / upsert_batch (with end_time)

    Note over S: dict entry gone — no leak

Loading

_{Reviews (2): Last reviewed commit: "fix(tests): Use context manager patches ..." | Re-trigger Greptile}