perf(api): Stream JSON query responses to reduce peak memory

[phacops]

Context

snuba-api has been OOMing in production. While auditing the read path for memory hot spots, one clear, contained win stood out: every query response is fully serialized into a single JSON string via json.dumps before being returned, so the process holds that entire string in memory in addition to the result object graph. Under concurrency this adds a full serialized-size copy per in-flight request.

(Worker-recycling guardrails API_WORKERS_MAX_RSS / API_WORKERS_LIFETIME are already set in prod, so the remaining lever is reducing per-request peak.)

Change

• Add stream_payload(), which encodes the response incrementally with simplejson's iterencode and yields chunks. The two query response callsites (/query, /snql, /mql via dataset_query, plus the storage-delete result path) now return this generator directly. Werkzeug treats it as a streamed response and Granian writes chunks as they are produced, so we never hold the whole serialized body at once.
• Bytes handling moved inline into a custom default (_payload_default) with encoding=None. This is required for streaming: simplejson otherwise raises UnicodeDecodeError on invalid-UTF8 bytes, which on a streamed response would fire mid-stream after bytes were already sent. The inline handler decodes valid UTF-8 and hex-encodes the rest, preserving the existing RAW_BYTESTRING__<hex> behavior. This replaces the old post-hoc _sanitize_payload retry.
• dump_payload() is kept as a thin wrapper ("".join(stream_payload(...))) so existing callers/tests are unaffected.

Behavioral note

Streamed responses use chunked transfer encoding and carry no Content-Length header (previously computed from the full string). The primary consumer (sentry's snuba client over urllib3) handles chunked responses fine.

Out of scope (follow-ups)

This only removes the serialized-string copy. The other copies on the read path — the native driver's buffered rows, the list[dict] rebuild in clickhouse/native.py, and the result-cache rapidjson.dumps copy — require rearchitecting the Reader/cache contract and are intentionally left for separate work.

Testing

• pytest tests/web/test_views.py — existing test_response_dumping unchanged; added test_stream_payload (incremental chunks + valid/invalid bytes) and test_streamed_response (generator Response is is_streamed and parses end-to-end).
• mypy snuba/web/views.py and pre-commit clean.

🤖 Generated with Claude Code

Agent transcript: https://claudescope.sentry.dev/share/8nrIX8NvceNRXqVEO5-f1DnOqTRtDlATlFL7Z04Z9ts