mirror of
https://github.com/headroomlabs-ai/headroom.git
synced 2026-11-22 13:29:19 +08:00
fix(metrics): cap savings_by_source rows and fold the rest into other (#4257)
## Description ### What we found `PrometheusMetrics.savings_by_source` keeps one row per `(source, realized)` for the life of the process, and `/metrics` exports three series per row (`headroom_savings_attribution_events_total`, `headroom_savings_attributed_tokens_total`, `headroom_savings_attributed_usd_total`). Source names come from extensions and turn hooks (`ctx.record_savings`, `record_scope_savings`). A plugin that puts something per-request into a source name would grow the table, and the exported series, without bound. @JerrettDavis raised this in the review of #4233. #4233 bounds the per-request ledger (256 named rows plus `other`). This PR applies the same bound to the process-wide table, which #4233 deliberately left alone. ## Type of Change - [x] Bug fix (non-breaking change that fixes an issue) ## Changes Made - `MAX_DISTINCT_SAVINGS_ROWS = 256` in `headroom/proxy/prometheus_metrics.py`, next to `MAX_DISTINCT_PATHS`. It follows the same pattern as the model and path caps in that file: - an existing row keeps accumulating; - past the cap, a new source is summed into an `other` row with the same `realized` flag; - one warning is logged the first time, and `reset_runtime` re-arms it. - As in the per-request ledger, `tool_search` never folds here (shared `UNFOLDED_SOURCES`): the documented headline query selects `source="tool_search"`, so a tool-search saving first seen after the cap keeps its named row. - Only the source name is lost. Every token, dollar and event is still booked. Every reader of `savings_by_source` (`cost.py`'s realized/projected USD sums, `/stats` `savings.by_source`, the `/metrics` export) adds rows together, so their totals are unchanged. ### Before / after | 306 distinct sources, plus `tool_search` on every request | origin/main | this PR | |---|---|---| | `savings_by_source` rows | 307 | 257 (255 named + `tool_search` + `other`) | | exported `headroom_savings_attributed_tokens_total` series | 307 | 257 | | tokens / USD / events booked | all | all | ## Testing - [x] Unit tests pass (`pytest`) - [x] Linting passes (`ruff check .`) - [x] Type checking passes (`mypy headroom`) - [x] New tests added for new functionality - [x] Manual testing performed - `tests/test_observability_metrics.py::test_prometheus_metrics_caps_savings_source_cardinality`: 306 requests each book a new realized source plus an estimated `tool_search` row through `record_request`. The test asserts 257 rows, 257 exported series, `tool_search` keeps its own row, and tokens, USD and events are conserved. It also checks that `reset_runtime` re-arms the warning. With the cap disabled it fails with `assert 307 == (256 + 1)`. - `tests/test_observability_metrics.py::test_tool_search_first_seen_after_the_savings_cap_keeps_its_row[True|False]`: 256 plugin rows fill the table, then `tool_search` and another new source arrive. `tool_search` keeps its own row and exported series with all 100 tokens, while the other new source goes to `other`. Both cases fail before7196e28d. - Known edge, past 256 sources only: once #4233 lands, a source folded here that also calls `record_extension_savings` appears in `extension_savings` both under its name and inside `other`. - Related suites pass: `test_observability_metrics.py`, `test_extension_attribution.py`, `test_compression_observability.py`, `test_turn_hooks.py` (78 passed). - `ruff check`, `ruff format --check` and `mypy` pass on the changed files. ### Test Output ```text # cap disabled: E AssertionError: assert 307 == (256 + 1) # with the cap: 78 passed in 10.51s ``` ## Real Behavior Proof - Environment: macOS 26.7, Python 3.12, this branch running `headroom proxy` on 127.0.0.1:8799 with an isolated HOME. Claude Code 2.1.302 (`claude-sonnet-5-5`) and Codex CLI 0.153.2 (`gpt-5.6-terra`, Responses over WebSocket) ran through it with real API keys. - Exact command / steps: both clients ran the same prompt ("read each .py file in this directory one at a time, then list each file's top-level functions") over three source files. Then `GET /stats` and `GET /metrics`. - Observed result: both clients completed (Claude 5 requests, Codex 9). `/stats` `savings.by_source` and `/metrics` show the `tool_search` row (5 events, 70,315 tokens) exactly as on main. No cap warning was logged, because real traffic stays far below the cap. - Not tested: a real plugin that mints unbounded source names. The cap path is covered by the unit test. ## Runtime Rollout Safety - Rollout-managed feature(s): none - Minimum rollout channel: n/a - Stable/default behavior changed: only past 256 distinct savings rows in one process, where new sources report as `other` - Kill switch / disable path: none needed; reporting only - Unsafe override required: no - Qualification impact: none - Rollback path: revert this PR ## Review Readiness - [x] I have performed a self-review - plus the code-reviewer agent - [x] This PR is ready for human review ## Additional Notes - This PR touches a different hunk of `record_request` from #4233, which changes `+ 1` to `+ event_count(item)` a few lines below. The two merge cleanly in either order. 🤖 Generated with [Claude Code](https://claude.com/claude-code) <!-- devin-review-badge-begin --> --- <a href="https://app.devin.ai/review/headroomlabs-ai/headroom/pull/4257" target="_blank"><picture><source media="(prefers-color-scheme: dark)" srcset="https://static.devin.ai/assets/gh-devin-review-dark.svg?v=4"><img src="https://static.devin.ai/assets/gh-devin-review-light.svg?v=4" alt="Devin Review"></picture></a> <!-- devin-review-badge-end --> ## Maintainer post-merge qualification Published `883a1a1f6085ed84ef7bacff69bcbd9ffaa098d1` incorporates upstream main `e4a60ed72f6ef5b28569f8bfbe2163cfda32c948`. 20 observability tests pass; both realized/non-realized late tool_search probes still fold the semantic label into other after arbitrary-label saturation. Ruff/format/diff pass; clean main sync, cap implementation unchanged. Existing Changes Requested decision remains applicable; unchanged findings are not reposted. Scoped Windows probes use the real CLI configuration harness or actual metrics object, not live cloud-provider traffic. Contributor correctness logic was not rewritten; hosted checks are handled by maintainers. Nothing merged. ### Maintainer current-main requalification Exact head `50514a724dd98ce672717eeb31344c88c13aeb1e`, upstream main `ecbaf86d78cf52d185fbffa306157b9a4cc1b906`. Mechanically updated from upstream/mainecbaf86d78with a clean merge and no contributor cap-policy rewrite. Re-reviewed existing discussion and integration delta. All 30 existing observability/Gemini cases pass; Ruff0.17.3 lint/format and whitespace checks pass. Both existing late-arriving semantic-source probes still fail after filling the row budget with arbitrary plugin labels: tool_search:1 and tool_search:0 are absent, with their 100 tokens folded into the respective other row. The probe verifies aggregate tokens are retained before asserting source identity, so this is the documented semantic-source/consumer problem rather than total-token loss. Retain the prior requested bounded tool_search source independently of arbitrary labels, cover both realized flags after saturation, and correct the claim that no consumer reads this table by source. This is not an inherited-main CI failure. Maintainers handle hosted checks; no contributor CI execution requested. Nothing merged. Exact-head hosted checks are being handled by maintainers. Existing correctness/qualification holds remain; nothing merged. ### Maintainer current-main requalification Exact head `20d7c9e9f00d83f7a5ec23dfbb2c30af6fa6aa06`, upstream main `8d3e9e5ca7cd645eb73084b848edbea5fc796e1a`. Clean mechanical sync with upstream main8d3e9e5ca7cd645eb73084b848edbea5fc796e1a with no contributor cap policy rewrite. Current discussion/source reviewed.30observability/Gemini tests pass;2changed Python files Ruff0.17.3check/format and whitespace pass. Both late-source saturation probes freshly fail: after256arbitrary plugin rows, late tool_search is absent for realized=true and false and its100tokens fold into other. The aggregate356tokens survives, but semantic source identity does not. Preserve important source identity under saturation and qualify actual overflow behavior; existing aggregate-only tests do not establish that contract. No inherited-main pipeline defect is alleged. Copied extension supports Python startup, not a new native build. Maintainers handle hosted checks; nothing merged. Exact-head hosted checks are being handled by maintainers. Existing correctness/qualification holds remain; nothing merged. ### Maintainer current-main requalification Exact head `4decdc96f1fa14fc8389bffd586ed15072a7b6fa`, upstream main `2d3ee51846435ef5b320eea4806af14e602b0d8c`. Clean integration with upstream/main2d3ee51846. Re-reviewed current discussion and composed metrics code. Exact comparison of prometheus_metrics.py and its observability tests against prior reviewed20d7c9e9fshows no changes. All 107 observability/release/version/MCP/memory/Gemini tests pass; whitespace passes. Earlier changed-file lint/format remains scoped to identical inputs. Existing source-identity saturation blocker remains: a late tool_search source after 256 arbitrary plugin rows folds its 100 tokens into other for either realized value. Total accounting survives, but important source identity does not. The previous probes cover identical metrics inputs and were not relabeled as rerun. Preserve the existing request to retain important identities under overflow and test actual saturation behavior. No contributor cap policy rewritten or inherited-main CI defect attributed. Existing native extension used for Python startup, not a fresh native build. Maintainers handle checks; nothing merged. Exact-head hosted checks are being handled by maintainers. Existing correctness/qualification holds remain; nothing merged. ### Maintainer current-main requalification Exact head `d8f67f36052d270489a4ff26b1b36ab112d2afa3`, upstream main `c01cf07d943d816cfd9d97e6838ed4e38f1ff9a4`. Clean integration with upstream/mainc01cf07d94without changing contributor cap/allocation policy. Re-reviewed current author discussion, metrics consumers and Devin findings. All20 observability tests pass with the fresh current-main native binding; Ruff and whitespace pass. The existing late-source identity blocker freshly reproduces for both realized=True and realized=False: after256 arbitrary plugin rows, a first tool_search attribution of100 tokens becomes other. Aggregate tokens remain correct, but no tool_search named row exists. The documented headline query in docs/metrics-technical-guide.md still filters source="tool_search", so it omits this saving. Please preserve required semantic identities under saturation (or provide an equivalent bounded headline counter), and cover first-seen-late realized/projected tool-search events in regression/runtime proof. Numeric conservation alone does not prove the documented metric contract. The OpenTelemetry implementation is byte-identical to main; its existing source-label behavior is not added as a PR-specific approval hold. No contributor correctness logic rewritten, no live collector/native-rebuild claim and no inherited-main pipeline defect attributed. Maintainers handle hosted checks; nothing merged. Exact-head hosted checks are being handled by maintainers. Existing correctness/qualification holds remain; nothing merged. ### Maintainer current-main requalification Exact head `d9c47a5521080675c4f6e0a0a3dc66fb8c72d79c`, upstream main `da6d9736cd82074988c7d0a2360b49d9d8b6c0a0`. Current-main requalification againstda6d9736cd. All69 integration patches exactly match c01cf07-to-current-main delta after diff-location normalization; contributor cap/allocation policy unchanged.20 observability tests pass.31 changed Python files pass Ruff lint/format; offline lock and whitespace pass. Existing pr3982 native DLL verified complete native input hashes/artifact checksum; no fresh build or live collector claim. Author/Devin discussions and metric consumers re-reviewed. Existing late-source identity blocker freshly reproduces for realized=True and False: after256 plugin rows, first tool_search100tokens becomes other; aggregate conserved but named tool_search row absent. docs/metrics-technical-guide.md headline query filters source=tool_search and omits that saving. Preserve required semantic identities under saturation or supply equivalent bounded headline counter; cover first-seen-late realized/projected tool_search events. Main-identical OpenTelemetry behavior is not a PR-specific hold. Maintainers handle hosted checks; no contributor correctness logic rewritten or merge. Exact-head hosted checks are being handled by maintainers. Existing correctness/qualification holds remain; nothing merged. ### Maintainer current-main requalification Exact head `fa88cf4bcab5e3db9448a1b3cc5f628b07a81a48`, upstream main `0ddaa56c21cdc01c4774391fdcc69e987ecc4e5c`. Clean integration with upstream/main0ddaa56c21. All 19 integration patches exactly match the prior-reviewed-mainda6d9736to current-main delta after diff-location normalization; contributor cap/allocation policy is unchanged. Fresh observability, doctor and incoming Vertex/backend/CLI suite: 309 passed, 11 skipped, 1 warning. Fourteen changed Python files pass Ruff lint/format; offline lock and whitespace checks pass. Author/Devin discussion and metric-consumer contract reviewed. Both synthetic saturation probes freshly reproduce the existing late-source identity finding: after 256 plugin rows, first tool_search 100 tokens folds into other for realized=True and False. Aggregate tokens are conserved, but the documented source=tool_search query misses that saving. Preserve required semantic identity under saturation or provide an equivalent bounded headline counter, with first-seen-late realized/projected regressions. No live collector behavior claimed. Static production/test diff is scoped to metrics aggregation and synthetic in-memory coverage; no new workflow/dependency/setup path. The inherited flagged workflow-validation job is excluded by unchanged workflow paths. Complete native input set/hashes and artifact checksum verified before DLL reuse; no fresh build claimed. Main-identical OpenTelemetry behavior is not attributed to this PR. Maintainers handle exact-head checks; no PR merged. Exact-head hosted checks are being handled by maintainers. Existing correctness/qualification holds remain; nothing merged. --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: JerrettDavis <mxjerrett@gmail.com>
This commit is contained in:
co-authored by
Claude Opus 5.5
JerrettDavis
parent
eb9932a671
commit
2038fb2a36
@@ -22,7 +22,7 @@ if TYPE_CHECKING:
|
||||
|
||||
from headroom import savings_ledger
|
||||
from headroom.observability import get_otel_metrics
|
||||
from headroom.proxy.savings_attribution import event_count
|
||||
from headroom.proxy.savings_attribution import UNFOLDED_SOURCES, event_count
|
||||
from headroom.proxy.savings_tracker import SavingsTracker, estimate_request_savings_usd
|
||||
|
||||
logger = logging.getLogger("headroom.proxy")
|
||||
@@ -40,6 +40,15 @@ _OTHER_MODEL = "other"
|
||||
MAX_DISTINCT_PATHS = 256
|
||||
_OTHER_PATH = "other"
|
||||
|
||||
# Same idea for savings attribution rows (see record_request). Source names come
|
||||
# from extensions and turn hooks, so one that mints a name per request would grow
|
||||
# savings_by_source, and its three exported series, without bound. Past this many
|
||||
# rows, new sources are summed into ``_OTHER_SOURCE``: only the name is lost,
|
||||
# never a token or a dollar. As in the per-request ledger, ``UNFOLDED_SOURCES``
|
||||
# never fold: the documented headline query selects ``source="tool_search"``.
|
||||
MAX_DISTINCT_SAVINGS_ROWS = 256
|
||||
_OTHER_SOURCE = "other"
|
||||
|
||||
# Closed label set for headroom_requests_rate_limited_total{source}. Two 429s
|
||||
# mean opposite things to an operator: "headroom" is OUR limiter refusing the
|
||||
# request (raise the cap), "upstream" is the provider refusing it (back off or
|
||||
@@ -209,6 +218,7 @@ class PrometheusMetrics:
|
||||
self.extension_savings: dict[str, int] = defaultdict(int)
|
||||
# Named savings attribution; realized and projected rows stay separate.
|
||||
self.savings_by_source: dict[str, dict[str, str | int | float | bool]] = {}
|
||||
self._savings_source_cardinality_warned = False
|
||||
|
||||
# Fail-open compression failures, keyed by reason ("timeout",
|
||||
# "error"). The proxy fails open on any optimization error so the
|
||||
@@ -449,6 +459,7 @@ class PrometheusMetrics:
|
||||
self.tokens_saved_by_strategy.clear()
|
||||
self.extension_savings.clear()
|
||||
self.savings_by_source.clear()
|
||||
self._savings_source_cardinality_warned = False
|
||||
with self._obs_counter_lock:
|
||||
self.compression_failed_by_reason.clear()
|
||||
self.upstream_connection_errors_by_provider.clear()
|
||||
@@ -984,6 +995,19 @@ class PrometheusMetrics:
|
||||
source = str(item.get("source") or "other")[:64]
|
||||
realized = bool(item.get("realized", True))
|
||||
key = f"{source}:{int(realized)}"
|
||||
if (
|
||||
key not in self.savings_by_source
|
||||
and source not in UNFOLDED_SOURCES
|
||||
and len(self.savings_by_source) >= MAX_DISTINCT_SAVINGS_ROWS
|
||||
):
|
||||
source, key = _OTHER_SOURCE, f"{_OTHER_SOURCE}:{int(realized)}"
|
||||
if not self._savings_source_cardinality_warned:
|
||||
self._savings_source_cardinality_warned = True
|
||||
logger.warning(
|
||||
"metrics.record: savings source cardinality cap (%d) reached; "
|
||||
'summing further sources into "other"',
|
||||
MAX_DISTINCT_SAVINGS_ROWS,
|
||||
)
|
||||
row = self.savings_by_source.setdefault(
|
||||
key,
|
||||
{
|
||||
|
||||
@@ -58,8 +58,9 @@ MAX_STAGE_MS = 3_600_000.0
|
||||
MAX_NAMED_ROWS = 256
|
||||
OVERFLOW_SOURCE = "other"
|
||||
# Core-booked sources never fold: ``tool_schema_savings_policy`` finds the
|
||||
# deferral's row by name to lower it when hooks un-defer tools.
|
||||
_UNFOLDED_SOURCES = frozenset({"tool_search"})
|
||||
# deferral's row by name to lower it when hooks un-defer tools, and the
|
||||
# metrics headline query selects ``source="tool_search"``.
|
||||
UNFOLDED_SOURCES = frozenset({"tool_search"})
|
||||
|
||||
# A row's sums saturate here instead of overflowing to ``inf`` (see
|
||||
# MAX_STAGE_MS for what a non-finite value does downstream). Both are far
|
||||
@@ -232,16 +233,12 @@ def accumulate(rows: list[dict[str, Any]], items: Iterable[dict[str, Any]]) -> N
|
||||
``tokens``, ``usd`` and ``events``; ``details`` stay only while every
|
||||
summed row agrees. Once ``rows`` holds ``MAX_NAMED_ROWS`` rows, a new
|
||||
source folds into ``OVERFLOW_SOURCE`` (which also takes unnamed savings),
|
||||
so the list grows by at most the overflow and ``_UNFOLDED_SOURCES`` keys.
|
||||
so the list grows by at most the overflow and ``UNFOLDED_SOURCES`` keys.
|
||||
"""
|
||||
for item in items:
|
||||
key = _key(item)
|
||||
target = _row_for(rows, key)
|
||||
if (
|
||||
target is None
|
||||
and key[0] not in _UNFOLDED_SOURCES
|
||||
and _row_count(rows) >= MAX_NAMED_ROWS
|
||||
):
|
||||
if target is None and key[0] not in UNFOLDED_SOURCES and _row_count(rows) >= MAX_NAMED_ROWS:
|
||||
if key[0] != OVERFLOW_SOURCE:
|
||||
# Details describe one source; the overflow row mixes several.
|
||||
item = {k: v for k, v in item.items() if k != "details"}
|
||||
|
||||
@@ -18,7 +18,11 @@ from headroom.observability import (
|
||||
set_otel_metrics,
|
||||
unregister_otel_metric_attribute_provider,
|
||||
)
|
||||
from headroom.proxy.prometheus_metrics import MAX_DISTINCT_PATHS, PrometheusMetrics
|
||||
from headroom.proxy.prometheus_metrics import (
|
||||
MAX_DISTINCT_PATHS,
|
||||
MAX_DISTINCT_SAVINGS_ROWS,
|
||||
PrometheusMetrics,
|
||||
)
|
||||
from headroom.telemetry.context import MAX_DISTINCT_MODELS
|
||||
from headroom.transforms.pipeline import TransformPipeline
|
||||
|
||||
@@ -624,6 +628,87 @@ async def test_prometheus_metrics_reset_rearms_path_cardinality_warning() -> Non
|
||||
assert len(metrics.inbound_requests_by_path) == 0
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_prometheus_metrics_caps_savings_source_cardinality() -> None:
|
||||
"""An extension that names a new savings source per request cannot grow
|
||||
savings_by_source, or its exported series, past MAX_DISTINCT_SAVINGS_ROWS plus
|
||||
one "other" row per realized flag, and every token and dollar is still booked."""
|
||||
metrics = PrometheusMetrics(stateless=True)
|
||||
sources = MAX_DISTINCT_SAVINGS_ROWS + 50
|
||||
for i in range(sources):
|
||||
await metrics.record_request(
|
||||
provider="openai",
|
||||
model="gpt-4o",
|
||||
input_tokens=10,
|
||||
output_tokens=1,
|
||||
tokens_saved=0,
|
||||
latency_ms=1.0,
|
||||
savings_attribution=[
|
||||
{"source": f"plugin_{i}", "realized": True, "tokens": 100, "usd": 0.01},
|
||||
{"source": "tool_search", "realized": False, "tokens": 10},
|
||||
],
|
||||
)
|
||||
|
||||
rows = metrics.savings_by_source
|
||||
assert len(rows) == MAX_DISTINCT_SAVINGS_ROWS + 1
|
||||
assert rows["tool_search:0"]["tokens"] == 10 * sources
|
||||
assert sum(int(r["tokens"]) for r in rows.values()) == 110 * sources
|
||||
assert sum(float(r["usd"]) for r in rows.values()) == pytest.approx(0.01 * sources)
|
||||
assert sum(int(r["events"]) for r in rows.values()) == 2 * sources
|
||||
exported = await metrics.export()
|
||||
series = [
|
||||
line
|
||||
for line in exported.splitlines()
|
||||
if line.startswith("headroom_savings_attributed_tokens_total{")
|
||||
]
|
||||
assert len(series) == MAX_DISTINCT_SAVINGS_ROWS + 1
|
||||
assert metrics._savings_source_cardinality_warned is True
|
||||
|
||||
await metrics.reset_runtime()
|
||||
|
||||
assert metrics._savings_source_cardinality_warned is False
|
||||
assert metrics.savings_by_source == {}
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
@pytest.mark.parametrize("realized", [True, False])
|
||||
async def test_tool_search_first_seen_after_the_savings_cap_keeps_its_row(realized) -> None:
|
||||
"""The headline query selects source="tool_search", so a tool-search saving
|
||||
that first arrives after the cap must keep its named row, not fold."""
|
||||
metrics = PrometheusMetrics(stateless=True)
|
||||
for i in range(MAX_DISTINCT_SAVINGS_ROWS):
|
||||
await metrics.record_request(
|
||||
provider="openai",
|
||||
model="gpt-4o",
|
||||
input_tokens=10,
|
||||
output_tokens=1,
|
||||
tokens_saved=0,
|
||||
latency_ms=1.0,
|
||||
savings_attribution=[{"source": f"plugin_{i}", "realized": realized, "tokens": 1}],
|
||||
)
|
||||
await metrics.record_request(
|
||||
provider="openai",
|
||||
model="gpt-4o",
|
||||
input_tokens=10,
|
||||
output_tokens=1,
|
||||
tokens_saved=0,
|
||||
latency_ms=1.0,
|
||||
savings_attribution=[
|
||||
{"source": "tool_search", "realized": realized, "tokens": 100},
|
||||
{"source": "late_plugin", "realized": realized, "tokens": 7},
|
||||
],
|
||||
)
|
||||
|
||||
rows = metrics.savings_by_source
|
||||
assert rows[f"tool_search:{int(realized)}"]["tokens"] == 100
|
||||
assert rows[f"other:{int(realized)}"]["tokens"] == 7
|
||||
exported = await metrics.export()
|
||||
assert (
|
||||
f'headroom_savings_attributed_tokens_total{{realized="{str(realized).lower()}",'
|
||||
f'source="tool_search"}} 100' in exported
|
||||
)
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_prometheus_metrics_export_bounds_model_series() -> None:
|
||||
"""export() emits at most MAX_DISTINCT_MODELS model series plus the 'other' bucket."""
|
||||
|
||||
Reference in New Issue
Block a user