GeezerVison

Summary

The Agentex tracing processor writes every span twice — a spans.create on start and a spans.update on end. The start row is overwritten by the end write moments later, so persisting it doubles per-span HTTP/DB write volume against the Agentex control plane. Under load this pressures the Agentex Postgres load.

This defaults the Agentex backend to end-only ingest: skip the start write and persist each span once, as a single spans.create on end.

Key correctness detail

Unlike the SGP processor (uniform upsert_batch), the Agentex processor does create on start + update on end. Naively skipping start would make the end update hit a non-existent row (404). So when the start is skipped, on_span_end now does a single spans.create carrying the complete span — one INSERT per span instead of INSERT + UPDATE. Applied to both the sync and async processors.

Scope / flag

• Gated by a dedicated, default-ON env var AGENTEX_TRACING_SKIP_AGENTEX_SPAN_START.
• Set to 0/false/no/off to restore the start write.

Trade-off

Same as the SGP end-only path: in-flight spans aren't visible until they complete, and spans whose process crashes before end won't persist. Documented in the docstring.

Tests

• pytest tests/lib/core/tracing/processors/test_agentex_tracing_processor.py — 7 passed (3 existing + 4 new: default-skip and skip-disabled, sync + async).
• pytest tests/lib/core/tracing — 85 passed, 2 skipped.
• ruff check clean.

Related

• Complements fix(tracing): fail open temporal span activities #437 (fail-open Temporal span activities) and the agent-side EGP-only export flag — together they remove the per-span Agentex HTTP load that was failing workflows under load.

🤖 Generated with Claude Code

Greptile Summary

This PR introduces an "end-only ingest" mode for the Agentex tracing processor, skipping the per-span spans.create call on start and converting the end-of-span write from spans.update to a single spans.create carrying the complete span — halving the HTTP/DB write volume against the Agentex control plane.

• Skip flag captured at init: both sync and async processors store self._skip_span_start once in __init__, ensuring the two halves of each span's lifecycle always agree on the mode (avoids the previously flagged 404/double-create split).
• _create_kwargs helper: extracts the full-span keyword dict shared by the start write (skip=OFF) and the end-only insert (skip=ON), removing the field-list duplication between sync and async paths.
• Test coverage: 4 new test cases verify default-skip, skip-disabled, and the mid-span toggle regression for both sync and async processors.

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    INIT["__init__(config)\nself._skip_span_start = _skip_span_start_enabled()"]

    INIT --> START["on_span_start(span)"]
    INIT --> END["on_span_end(span)"]

    START --> C1{self._skip_span_start?}
    C1 -- "True (default)" --> NOOP["return (no-op)"]
    C1 -- "False" --> CREATE_START["spans.create(**_create_kwargs(span))"]

    END --> C2{self._skip_span_start?}
    C2 -- "True (default)" --> CREATE_END["spans.create(**_create_kwargs(span))\n(single INSERT — complete span)"]
    C2 -- "False" --> UPDATE_END["spans.update(span.id, **span.model_dump(...))"]

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    INIT["__init__(config)\nself._skip_span_start = _skip_span_start_enabled()"]

    INIT --> START["on_span_start(span)"]
    INIT --> END["on_span_end(span)"]

    START --> C1{self._skip_span_start?}
    C1 -- "True (default)" --> NOOP["return (no-op)"]
    C1 -- "False" --> CREATE_START["spans.create(**_create_kwargs(span))"]

    END --> C2{self._skip_span_start?}
    C2 -- "True (default)" --> CREATE_END["spans.create(**_create_kwargs(span))\n(single INSERT — complete span)"]
    C2 -- "False" --> UPDATE_END["spans.update(span.id, **span.model_dump(...))"]

_{Reviews (3): Last reviewed commit: "fix(tracing): capture skip-span-start de..." | Re-trigger Greptile}