[codex] Add turn profiling analytics (#26484)

## Summary

Add flat profiling fields to `codex_turn_event` so analytics can explain
where turn wall-clock time is spent without changing tool execution
behavior.

The profile reports:
- time before the first sampling request
- sampling time across all attempts and follow-ups
- overhead between sampling requests
- time blocked in the post-sampling tool drain
- time after the final sampling request
- sampling request and retry counts

## Implementation

- Extend the existing turn timing state with constant-memory phase
accounting and one RAII phase guard.
- Observe sampling and the existing post-sampling drain only at turn
orchestration boundaries.
- Keep tool runtime, tool futures, response item handling, and turn
lifecycle values unchanged.
- Add the profiling fields directly to the existing analytics turn event
without changing app-server protocol or rollout persistence.
- Use the existing turn `status` to distinguish completed, failed, and
interrupted profiles.

Exact sampling/tool overlap is intentionally omitted because measuring
tool completion accurately would require hooks in the tool execution
path.

## Validation

- Add app-server end-to-end coverage for a single-sampling turn with no
blocking tool work.
- Add app-server end-to-end coverage for `request_user_input` blocking
followed by a second sampling request.
- CI is running on the PR; tests were not executed locally per
repository guidance.
This commit is contained in:
Ahmed Ibrahim
2026-06-05 11:27:10 -07:00
committed by GitHub
Unverified
parent 82b15b65e2
commit 8d72fb6de9
11 changed files with 530 additions and 101 deletions
+9
View File
@@ -1088,6 +1088,7 @@ async fn run_sampling_request(
ResponsesStreamRequest::Sampling,
)
.await?;
turn_context.turn_timing_state.record_sampling_retry();
}
}
@@ -1800,6 +1801,7 @@ async fn try_run_sampling_request(
turn_context.model_info.slug.as_str(),
turn_context.provider.info().name.as_str(),
);
let sampling_timing_guard = turn_context.turn_timing_state.begin_sampling();
let mut stream = client_session
.stream(
prompt,
@@ -2213,6 +2215,7 @@ async fn try_run_sampling_request(
}
}
};
drop(sampling_timing_guard);
flush_assistant_text_segments_all(
&sess,
@@ -2222,7 +2225,13 @@ async fn try_run_sampling_request(
)
.await;
let tool_blocking_timing_guard = if in_flight.is_empty() {
None
} else {
Some(turn_context.turn_timing_state.begin_tool_blocking())
};
drain_in_flight(&mut in_flight, sess.clone(), turn_context.clone()).await?;
drop(tool_blocking_timing_guard);
if should_emit_token_count {
// A tool call such as request_user_input can intentionally pause the turn. Emit token
+13
View File
@@ -33,6 +33,7 @@ use crate::session::turn_context::TurnContext;
use crate::state::ActiveTurn;
use crate::state::RunningTask;
use crate::state::TaskKind;
use codex_analytics::TurnProfileFact;
use codex_analytics::TurnTokenUsageFact;
use codex_login::AuthManager;
use codex_models_manager::manager::SharedModelsManager;
@@ -751,6 +752,12 @@ impl Session {
.turn_timing_state
.time_to_first_token_ms()
.await;
self.services
.analytics_events_client
.track_turn_profile(TurnProfileFact {
turn_id: turn_context.sub_id.clone(),
profile: turn_context.turn_timing_state.complete_profile(),
});
self.emit_turn_stop_lifecycle(turn_context.extension_data.as_ref())
.await;
if let Err(err) = self
@@ -868,6 +875,12 @@ impl Session {
.turn_timing_state
.completed_at_and_duration_ms()
.await;
self.services
.analytics_events_client
.track_turn_profile(TurnProfileFact {
turn_id: task.turn_context.sub_id.clone(),
profile: task.turn_context.turn_timing_state.complete_profile(),
});
let event = EventMsg::TurnAborted(TurnAbortedEvent {
turn_id: Some(task.turn_context.sub_id.clone()),
reason,
+191
View File
@@ -1,8 +1,11 @@
use std::sync::Arc;
use std::sync::Mutex as StdMutex;
use std::time::Duration;
use std::time::Instant;
use std::time::SystemTime;
use std::time::UNIX_EPOCH;
use codex_analytics::TurnProfile;
use codex_otel::TURN_TTFM_DURATION_METRIC;
use codex_protocol::items::TurnItem;
use codex_protocol::models::ResponseItem;
@@ -39,6 +42,7 @@ pub(crate) async fn record_turn_ttfm_metric(turn_context: &TurnContext, item: &T
#[derive(Debug, Default)]
pub(crate) struct TurnTimingState {
state: Mutex<TurnTimingStateInner>,
profile: StdMutex<TurnProfileState>,
}
#[derive(Debug, Default)]
@@ -49,6 +53,35 @@ struct TurnTimingStateInner {
first_message_at: Option<Instant>,
}
#[derive(Debug, Default)]
struct TurnProfileState {
started_at: Option<Instant>,
last_transition_at: Option<Instant>,
active_phase: Option<TurnProfilePhase>,
seen_sampling: bool,
before_first_sampling: Duration,
sampling: Duration,
between_sampling_overhead: Duration,
tool_blocking: Duration,
pending_idle_after_sampling: Duration,
sampling_request_count: u32,
sampling_retry_count: u32,
completed_profile: Option<TurnProfile>,
}
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
enum TurnProfilePhase {
Sampling,
ToolBlocking,
}
#[must_use]
pub(crate) struct TurnProfileTimingGuard {
timing: Arc<TurnTimingState>,
phase: TurnProfilePhase,
active: bool,
}
impl TurnTimingState {
pub(crate) async fn mark_turn_started(&self, started_at: Instant) -> i64 {
let started_at_unix_ms = now_unix_timestamp_ms();
@@ -57,6 +90,7 @@ impl TurnTimingState {
state.started_at_unix_secs = Some(started_at_unix_ms / 1000);
state.first_token_at = None;
state.first_message_at = None;
self.profile_state().start(started_at);
started_at_unix_ms
}
@@ -80,6 +114,32 @@ impl TurnTimingState {
.map(|duration| i64::try_from(duration.as_millis()).unwrap_or(i64::MAX))
}
pub(crate) fn complete_profile(&self) -> TurnProfile {
self.profile_state().complete(Instant::now())
}
pub(crate) fn begin_sampling(self: &Arc<Self>) -> TurnProfileTimingGuard {
let active = self.profile_state().begin_sampling(Instant::now());
TurnProfileTimingGuard {
timing: Arc::clone(self),
phase: TurnProfilePhase::Sampling,
active,
}
}
pub(crate) fn record_sampling_retry(&self) {
self.profile_state().record_sampling_retry();
}
pub(crate) fn begin_tool_blocking(self: &Arc<Self>) -> TurnProfileTimingGuard {
let active = self.profile_state().begin_tool_blocking(Instant::now());
TurnProfileTimingGuard {
timing: Arc::clone(self),
phase: TurnProfilePhase::ToolBlocking,
active,
}
}
pub(crate) async fn record_ttft_for_response_event(
&self,
event: &ResponseEvent,
@@ -98,6 +158,22 @@ impl TurnTimingState {
let mut state = self.state.lock().await;
state.record_turn_ttfm()
}
fn profile_state(&self) -> std::sync::MutexGuard<'_, TurnProfileState> {
self.profile
.lock()
.unwrap_or_else(std::sync::PoisonError::into_inner)
}
}
impl Drop for TurnProfileTimingGuard {
fn drop(&mut self) {
if self.active {
self.timing
.profile_state()
.end_phase(Instant::now(), self.phase);
}
}
}
fn now_unix_timestamp_secs() -> i64 {
@@ -111,6 +187,121 @@ pub(crate) fn now_unix_timestamp_ms() -> i64 {
i64::try_from(duration.as_millis()).unwrap_or(i64::MAX)
}
fn duration_to_u64_ms(duration: Duration) -> u64 {
u64::try_from(duration.as_millis()).unwrap_or(u64::MAX)
}
impl TurnProfileState {
fn start(&mut self, started_at: Instant) {
*self = Self {
started_at: Some(started_at),
last_transition_at: Some(started_at),
..Self::default()
};
}
fn begin_sampling(&mut self, now: Instant) -> bool {
if self.completed_profile.is_some()
|| self.started_at.is_none()
|| self.active_phase.is_some()
{
return false;
}
self.advance(now);
if self.seen_sampling {
self.between_sampling_overhead += std::mem::take(&mut self.pending_idle_after_sampling);
}
self.seen_sampling = true;
self.active_phase = Some(TurnProfilePhase::Sampling);
self.sampling_request_count = self.sampling_request_count.saturating_add(1);
true
}
fn record_sampling_retry(&mut self) {
if self.completed_profile.is_none() && self.started_at.is_some() {
self.sampling_retry_count = self.sampling_retry_count.saturating_add(1);
}
}
fn begin_tool_blocking(&mut self, now: Instant) -> bool {
if self.completed_profile.is_some()
|| self.started_at.is_none()
|| self.active_phase.is_some()
{
return false;
}
self.advance(now);
self.active_phase = Some(TurnProfilePhase::ToolBlocking);
true
}
fn end_phase(&mut self, now: Instant, phase: TurnProfilePhase) {
if self.completed_profile.is_some() || self.active_phase != Some(phase) {
return;
}
self.advance(now);
self.active_phase = None;
}
fn advance(&mut self, now: Instant) {
let Some(previous) = self.last_transition_at.replace(now) else {
return;
};
let elapsed = now.saturating_duration_since(previous);
match self.active_phase {
Some(TurnProfilePhase::Sampling) => self.sampling += elapsed,
Some(TurnProfilePhase::ToolBlocking) => self.tool_blocking += elapsed,
None if self.seen_sampling => self.pending_idle_after_sampling += elapsed,
None => self.before_first_sampling += elapsed,
}
}
fn complete(&mut self, now: Instant) -> TurnProfile {
if let Some(profile) = self.completed_profile.as_ref() {
return profile.clone();
}
let final_phase = self.active_phase;
self.advance(now);
let after_last_sampling = if self.seen_sampling {
std::mem::take(&mut self.pending_idle_after_sampling)
} else {
Duration::ZERO
};
let mut profile = TurnProfile {
before_first_sampling_ms: duration_to_u64_ms(self.before_first_sampling),
sampling_ms: duration_to_u64_ms(self.sampling),
between_sampling_overhead_ms: duration_to_u64_ms(self.between_sampling_overhead),
tool_blocking_ms: duration_to_u64_ms(self.tool_blocking),
after_last_sampling_ms: duration_to_u64_ms(after_last_sampling),
sampling_request_count: self.sampling_request_count,
sampling_retry_count: self.sampling_retry_count,
};
let total_ms = self
.started_at
.map(|started_at| duration_to_u64_ms(now.saturating_duration_since(started_at)))
.unwrap_or_default();
let classified_ms = profile
.before_first_sampling_ms
.saturating_add(profile.sampling_ms)
.saturating_add(profile.between_sampling_overhead_ms)
.saturating_add(profile.tool_blocking_ms)
.saturating_add(profile.after_last_sampling_ms);
let rounding_ms = total_ms.saturating_sub(classified_ms);
match final_phase {
Some(TurnProfilePhase::Sampling) => profile.sampling_ms += rounding_ms,
Some(TurnProfilePhase::ToolBlocking) => profile.tool_blocking_ms += rounding_ms,
None if self.seen_sampling => profile.after_last_sampling_ms += rounding_ms,
None => profile.before_first_sampling_ms += rounding_ms,
}
self.active_phase = None;
self.completed_profile = Some(profile.clone());
profile
}
}
impl TurnTimingStateInner {
fn time_to_first_token(&self) -> Option<Duration> {
Some(self.first_token_at?.duration_since(self.started_at?))
+41
View File
@@ -1,13 +1,17 @@
use codex_analytics::TurnProfile;
use codex_protocol::items::AgentMessageItem;
use codex_protocol::items::TurnItem;
use codex_protocol::models::ContentItem;
use codex_protocol::models::FunctionCallOutputPayload;
use codex_protocol::models::ResponseItem;
use pretty_assertions::assert_eq;
use std::time::Duration;
use std::time::Instant;
use std::time::SystemTime;
use std::time::UNIX_EPOCH;
use super::TurnProfilePhase;
use super::TurnProfileState;
use super::TurnTimingState;
use super::response_item_records_turn_ttft;
use crate::ResponseEvent;
@@ -146,3 +150,40 @@ fn response_item_records_turn_ttft_ignores_empty_non_output_items() {
}
));
}
#[test]
fn turn_profile_breaks_down_sampling_blocking_and_retry_overhead() {
let started_at = Instant::now();
let mut state = TurnProfileState::default();
state.start(started_at);
let _ = state.begin_sampling(started_at + Duration::from_millis(100));
state.end_phase(
started_at + Duration::from_millis(600),
TurnProfilePhase::Sampling,
);
let _ = state.begin_tool_blocking(started_at + Duration::from_millis(600));
state.end_phase(
started_at + Duration::from_millis(900),
TurnProfilePhase::ToolBlocking,
);
state.record_sampling_retry();
let _ = state.begin_sampling(started_at + Duration::from_millis(1_000));
state.end_phase(
started_at + Duration::from_millis(1_200),
TurnProfilePhase::Sampling,
);
assert_eq!(
state.complete(started_at + Duration::from_millis(1_300)),
TurnProfile {
before_first_sampling_ms: 100,
sampling_ms: 700,
between_sampling_overhead_ms: 100,
tool_blocking_ms: 300,
after_last_sampling_ms: 100,
sampling_request_count: 2,
sampling_retry_count: 1,
}
);
}