:hedgehog: PostHog is the leading platform for building self-driving products. Our developer tools – AI observability, analytics, session replay, flags, experiments, error tracking, logs, and more – capture all the context agents need to diagnose problems, uncover opportunities, and ship fixes. Steer it all from Slack, web, desktop, or the MCP.
Python ab-testingai-analyticsanalyticscdpdata-warehouseexperimentsfeature-flagsjavascriptproduct-analyticspythonreactsession-replaysurveystypescriptweb-analytics
perf(mcp-analytics): compute tool latency quantiles in one pass (#122813)
Paul D'Ambra9620ffd6a03f
diff --git a/products/mcp_analytics/backend/hogql_queries/tool_tables.py b/products/mcp_analytics/backend/hogql_queries/tool_tables.pyindex 49a89a8715b8..f065aa05ecaf 100644--- a/products/mcp_analytics/backend/hogql_queries/tool_tables.py+++ b/products/mcp_analytics/backend/hogql_queries/tool_tables.py@@ -428,8 +428,9 @@ class MCPToolFailureOccurrencesQueryRunner(AnalyticsQueryRunner[MCPToolFailureOc_IS_ERROR = "countIf(toBool(properties.$mcp_is_error))"-_P50 = "round(quantile(0.5)(toFloat(properties.$mcp_duration_ms)))"-_P95 = "round(quantile(0.95)(toFloat(properties.$mcp_duration_ms)))"+# One sample for both levels keeps p50 <= p95. quantiles() returns NaN, not NULL, when there are no+# durations, so map NaN back to NULL.+_P50_P95 = "arrayMap(q -> if(isNaN(q), NULL, round(q)), quantiles(0.5, 0.95)({duration_ms}))"class MCPToolStatsQueryRunner(AnalyticsQueryRunner[MCPToolStatsQueryResponse]):@@ -446,30 +447,34 @@ class MCPToolStatsQueryRunner(AnalyticsQueryRunner[MCPToolStatsQueryResponse]):def to_query(self) -> ast.SelectQuery | ast.SelectSetQuery:return parse_select("""- SELECT- countIf(is_tool) AS calls,- countIf(is_tool AND is_error) AS errors,- round(quantileIf(0.5)(duration_ms, is_tool)) AS p50_ms,- round(quantileIf(0.95)(duration_ms, is_tool)) AS p95_ms,- uniqIf(distinct_id, is_tool) AS users,- uniqIf(conversation_id, is_tool) AS conversations,- countIf(is_tool AND has_intent) AS with_intent,- count() AS total_calls,- uniq(conversation_id) AS total_conversations+ SELECT calls, errors, latency[1] AS p50_ms, latency[2] AS p95_ms, users, conversations, with_intent,+ total_calls, total_conversationsFROM (SELECT- distinct_id,- nullIf({_CONVERSATION_ID}, '') AS conversation_id,- toBool(properties.$mcp_is_error) AS is_error,- toFloat(properties.$mcp_duration_ms) AS duration_ms,- notEmpty(toString(properties.$mcp_intent)) AND toString(properties.$mcp_intent) != '{}' AS has_intent,- {is_tool} AS is_tool- FROM events- WHERE {where}+ countIf(is_tool) AS calls,+ countIf(is_tool AND is_error) AS errors,+ {_P50_P95} AS latency,+ uniqIf(distinct_id, is_tool) AS users,+ uniqIf(conversation_id, is_tool) AS conversations,+ countIf(is_tool AND has_intent) AS with_intent,+ count() AS total_calls,+ uniq(conversation_id) AS total_conversations+ FROM (+ SELECT+ distinct_id,+ nullIf({_CONVERSATION_ID}, '') AS conversation_id,+ toBool(properties.$mcp_is_error) AS is_error,+ toFloat(properties.$mcp_duration_ms) AS duration_ms,+ notEmpty(toString(properties.$mcp_intent)) AND toString(properties.$mcp_intent) != '{}' AS has_intent,+ {is_tool} AS is_tool+ FROM events+ WHERE {where}+ ))""",placeholders={"_CONVERSATION_ID": parse_expr(CONVERSATION_ID_SQL),+ "_P50_P95": parse_expr(_P50_P95, {"duration_ms": parse_expr("if(is_tool, duration_ms, NULL)")}),"is_tool": effective_tool_expr(self.query.toolName),"where": _tool_call_where(self.query, self.query_date_range, self.team, scope_to_tool=False),},@@ -538,17 +543,19 @@ class MCPToolDailyStatsQueryRunner(AnalyticsQueryRunner[MCPToolDailyStatsQueryReinterval = self.query.interval.value if self.query.interval else "day"return parse_select("""- SELECT- toString(dateTrunc({interval}, timestamp)) AS day,- count() AS calls,- {_IS_ERROR} AS errors,- {_P50} AS p50,- {_P95} AS p95,- uniq(distinct_id) AS users,- uniq(nullIf({_CONVERSATION_ID}, '')) AS sessions- FROM events- WHERE {where}- GROUP BY day+ SELECT day, calls, errors, latency[1] AS p50, latency[2] AS p95, users, sessions+ FROM (+ SELECT+ toString(dateTrunc({interval}, timestamp)) AS day,+ count() AS calls,+ {_IS_ERROR} AS errors,+ {_P50_P95} AS latency,+ uniq(distinct_id) AS users,+ uniq(nullIf({_CONVERSATION_ID}, '')) AS sessions+ FROM events+ WHERE {where}+ GROUP BY day+ )ORDER BY dayLIMIT 10000""",@@ -556,8 +563,7 @@ class MCPToolDailyStatsQueryRunner(AnalyticsQueryRunner[MCPToolDailyStatsQueryRe"interval": ast.Constant(value=interval),"_CONVERSATION_ID": parse_expr(CONVERSATION_ID_SQL),"_IS_ERROR": parse_expr(_IS_ERROR),- "_P50": parse_expr(_P50),- "_P95": parse_expr(_P95),+ "_P50_P95": parse_expr(_P50_P95, {"duration_ms": parse_expr("toFloat(properties.$mcp_duration_ms)")}),"where": _tool_call_where(self.query, self.query_date_range, self.team),},)diff --git a/products/mcp_analytics/backend/tests/test_tool_tables.py b/products/mcp_analytics/backend/tests/test_tool_tables.pyindex 8653af3115d3..4c2c3686fdc1 100644--- a/products/mcp_analytics/backend/tests/test_tool_tables.py+++ b/products/mcp_analytics/backend/tests/test_tool_tables.py@@ -502,7 +502,15 @@ class TestMCPToolStatsQueryRunner(_MCPAnalyticsTeamScopedTestMixin, ClickhouseTeassert row.conversations == 2# Only the '{"goal":"x"}' call counts; '{}' and missing do not.assert row.with_intent == 1- assert row.p50_ms is not None and row.p95_ms is not None+ assert (row.p50_ms, row.p95_ms) == (200, 290)++ def test_returns_none_latency_when_durations_are_missing(self) -> None:+ _emit_tool_call(self.team)+ flush_persons_and_events()++ rows = self._run()++ assert (rows[0].p50_ms, rows[0].p95_ms) == (None, None)def test_excludes_events_without_new_sdk_source(self) -> None:_emit_tool_call(self.team, source=None, duration_ms=100)@@ -566,8 +574,8 @@ class TestMCPToolDailyStatsQueryRunner(_MCPAnalyticsTeamScopedTestMixin, Clickhodef test_groups_by_day_in_order(self) -> None:now = datetime.now(tz=UTC)_emit_tool_call(self.team, distinct_id="d1", session_id="s1", timestamp=now - timedelta(days=2))- _emit_tool_call(self.team, distinct_id="d2", session_id="s2", timestamp=now)- _emit_tool_call(self.team, distinct_id="d3", session_id="s3", timestamp=now)+ _emit_tool_call(self.team, distinct_id="d2", session_id="s2", timestamp=now, duration_ms=100)+ _emit_tool_call(self.team, distinct_id="d3", session_id="s3", timestamp=now, duration_ms=300)flush_persons_and_events()rows = self._run()@@ -577,6 +585,8 @@ class TestMCPToolDailyStatsQueryRunner(_MCPAnalyticsTeamScopedTestMixin, Clickhoassert rows[0].calls == 1assert rows[1].calls == 2assert rows[1].sessions == 2+ assert (rows[0].p50, rows[0].p95) == (0, 0)+ assert (rows[1].p50, rows[1].p95) == (200, 290)def test_buckets_by_hour_when_interval_is_hour(self) -> None:# Two calls in the same day but different hours split into two hourly buckets — a sub-day