Files
sorsfonalak/services/worker/src/analysis/mod.rs
T
fecoandClaude Opus 5.5 bc0227042b
CI / test (push) Successful in 1m8s
CI / build (push) Successful in 4m9s
CI / deploy (push) Skipped
Worker: separate quick lane for summaries and answers
Summaries and questions get their own slots (QUICK_CONCURRENCY, default 2),
so they never wait behind detailed analyses (ANALYSIS_CONCURRENCY, default
4). A free slow slot also takes quick work; a lane of size 0 is off, so the
lanes can run as separate deployments. Documented that several workers can
share the queue.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012vLHzh5qUhFtmJhzkcTbXL
2026-11-19 06:40:09 +00:00

661 lines
26 KiB
Rust

//! Writes the requested text analyses (`analyses` rows in `pending` state).
//!
//! The backend inserts the row and charges the credits; this job claims it
//! (`running`), builds the prompt from the stored chart data, asks the model
//! through OpenRouter and stores the structured answer (`done`). Temporary
//! failures are retried with backoff; after the last attempt, or on a
//! permanent error, the analysis is `failed` and the unit is refunded.
//!
//! Admin test runs of an edited prompt (`prompt_previews`) go through the
//! same loop when no analysis is due: one attempt, no credits, no retries.
//!
//! Two lanes, each with its own slots, so a short text never waits behind
//! the long ones: the quick lane writes summaries and answers questions, the
//! slow lane writes the detailed analyses and forecasts (and the test runs).
//! A free slow slot also takes quick work. A lane of size 0 is off, so the
//! lanes can run as separate deployments; several workers may share the
//! queue, rows are claimed with `FOR UPDATE SKIP LOCKED`.
//!
//! Fallback: when the model chosen on the admin page fails twice (or once with
//! a permanent error), the default model (`fallback_model`) takes over with a
//! fresh set of attempts; only when that fails too is the unit refunded.
pub mod compact;
pub mod openrouter;
pub mod prompt;
pub mod questions;
use std::collections::HashSet;
use std::sync::{Arc, Mutex};
use std::time::{Duration, Instant};
use serde_json::Value;
use sqlx::PgPool;
use tokio::sync::{Semaphore, watch};
use uuid::Uuid;
use self::openrouter::{Completion, LlmError, OpenRouter};
use self::prompt::{Kind, PROMPT_VERSION, Prompt};
use self::questions::QuestionJob;
/// A `running` analysis older than this was left behind by a stopped worker.
const STUCK_AFTER_MINUTES: i32 = 15;
#[derive(Debug, sqlx::FromRow)]
pub struct Job {
pub id: Uuid,
pub kind: String,
pub chart_id: Option<Uuid>,
pub synastry_id: Option<Uuid>,
pub forecast_id: Option<Uuid>,
pub language: String,
pub model: String,
pub prompt_version: String,
pub attempts: i32,
/// The default model, when the chosen one is another (see the module docs).
pub fallback_model: Option<String>,
pub on_fallback: bool,
/// When this attempt started; also its token: an admin restart clears it,
/// so a superseded attempt neither writes its result nor keeps running.
pub started_at: time::OffsetDateTime,
}
impl Job {
/// The model of this attempt.
pub fn active_model(&self) -> &str {
active_model(&self.model, self.fallback_model.as_deref(), self.on_fallback)
}
}
pub fn active_model<'a>(model: &'a str, fallback: Option<&'a str>, on_fallback: bool) -> &'a str {
match fallback {
Some(f) if on_fallback => f,
_ => model,
}
}
/// What happens after a failed attempt.
#[derive(Debug, PartialEq, Eq)]
pub enum Next {
Retry(Duration),
/// Switch to the fallback model, with a fresh set of attempts, at once.
Fallback,
/// Give up (and refund).
Fail,
}
/// The chosen model gets two attempts (one after a permanent error), then the
/// fallback model takes over with its own `max_attempts`.
pub fn next_step(error: &LlmError, attempts: i32, max_attempts: i32, can_fall_back: bool) -> Next {
if can_fall_back && (matches!(error, LlmError::Permanent(_)) || attempts >= 2) {
return Next::Fallback;
}
match error {
LlmError::Transient(_) if attempts < max_attempts => Next::Retry(backoff(attempts)),
_ => Next::Fail,
}
}
/// A test run of an edited system prompt on a chart.
#[derive(Debug, sqlx::FromRow)]
pub struct PreviewJob {
pub id: Uuid,
pub kind: String,
pub chart_id: Option<Uuid>,
pub synastry_id: Option<Uuid>,
pub forecast_id: Option<Uuid>,
pub language: String,
pub model: String,
pub system: String,
}
/// Which analyses a claim may take.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Lane {
/// The short summary of a chart (questions are claimed separately).
Quick,
/// The detailed analyses and the yearly forecast.
Slow,
}
impl Lane {
fn kinds(self) -> &'static [&'static str] {
match self {
Lane::Quick => &["summary"],
Lane::Slow => &["natal", "synastry", "forecast"],
}
}
}
/// Slots per lane; 0 turns a lane off.
#[derive(Debug, Clone, Copy)]
pub struct Lanes {
pub quick: usize,
pub slow: usize,
}
pub struct Analyzer {
pub db: PgPool,
pub llm: OpenRouter,
pub max_attempts: i32,
pub lanes: Lanes,
/// Analyses and previews being written right now (released on shutdown).
running: Mutex<HashSet<Uuid>>,
}
/// Wait before the next attempt: 30 s, 2 min, 8 min, ... at most 1 hour.
pub fn backoff(attempts: i32) -> Duration {
let exp = attempts.clamp(1, 6).saturating_sub(1) as u32;
Duration::from_secs((30 * 4u64.pow(exp)).min(3600))
}
impl Analyzer {
pub fn new(db: PgPool, llm: OpenRouter, max_attempts: i32, lanes: Lanes) -> Self {
Self { db, llm, max_attempts, lanes, running: Mutex::default() }
}
/// Takes the next due analysis of the lane, if any.
pub async fn claim(&self, lane: Lane) -> Result<Option<Job>, sqlx::Error> {
let job: Option<Job> = sqlx::query_as(
"UPDATE analyses SET status = 'running', started_at = now(), attempts = attempts + 1 \
WHERE id = (SELECT id FROM analyses WHERE status = 'pending' AND next_attempt_at <= now() \
AND kind = ANY($1) \
ORDER BY next_attempt_at LIMIT 1 FOR UPDATE SKIP LOCKED) \
RETURNING id, kind, chart_id, synastry_id, forecast_id, language, model, prompt_version, attempts, \
fallback_model, on_fallback, started_at",
)
.bind(lane.kinds())
.fetch_optional(&self.db)
.await?;
if let Some(job) = &job {
self.running.lock().unwrap().insert(job.id);
}
Ok(job)
}
/// Builds the prompt from the stored chart or synastry.
pub async fn prompt_for(&self, job: &Job) -> Result<Prompt, LlmError> {
let kind = Kind::parse(&job.kind).ok_or_else(|| LlmError::Permanent(format!("unknown kind '{}'", job.kind)))?;
let db_err = |e: sqlx::Error| LlmError::Transient(format!("database: {e}"));
let text = match kind {
Kind::Natal | Kind::Summary => self.natal_text(job.chart_id).await?,
Kind::Synastry => self.synastry_text(job.synastry_id).await?,
Kind::Forecast => self.forecast_text(job.forecast_id).await?,
Kind::Question => return Err(LlmError::Permanent("a question is not an analysis".into())),
};
let (base, revision) = prompt::split_version(&job.prompt_version);
if base != PROMPT_VERSION {
tracing::warn!(id = %job.id, requested = %job.prompt_version, current = PROMPT_VERSION, "prompt version differs");
}
// An admin-edited system prompt; NULL (or a missing row) means the built-in one.
let system: Option<String> = match revision {
Some(id) => sqlx::query_scalar("SELECT system FROM prompt_revisions WHERE id = $1")
.bind(id)
.fetch_optional(&self.db)
.await
.map_err(db_err)?
.flatten(),
None => None,
};
Ok(prompt::build(kind, &job.language, &text, system.as_deref()))
}
/// The compact text of a stored natal chart.
async fn natal_text(&self, chart_id: Option<Uuid>) -> Result<String, LlmError> {
let (chart,): (Value,) = sqlx::query_as("SELECT result FROM charts WHERE id = $1")
.bind(chart_id)
.fetch_optional(&self.db)
.await
.map_err(|e| LlmError::Transient(format!("database: {e}")))?
.ok_or_else(|| LlmError::Permanent("the chart was deleted".into()))?;
Ok(compact::natal(&chart))
}
/// The compact text of a stored synastry with its two charts.
async fn synastry_text(&self, synastry_id: Option<Uuid>) -> Result<String, LlmError> {
let (result, chart1, chart2): (Value, Value, Value) = sqlx::query_as(
"SELECT s.result, c1.result, c2.result FROM synastries s \
JOIN charts c1 ON c1.id = s.chart1_id JOIN charts c2 ON c2.id = s.chart2_id WHERE s.id = $1",
)
.bind(synastry_id)
.fetch_optional(&self.db)
.await
.map_err(|e| LlmError::Transient(format!("database: {e}")))?
.ok_or_else(|| LlmError::Permanent("the synastry was deleted".into()))?;
Ok(compact::synastry(&chart1, &chart2, &result))
}
/// The compact text of a stored forecast with its natal chart.
async fn forecast_text(&self, forecast_id: Option<Uuid>) -> Result<String, LlmError> {
let (result, chart): (Value, Value) = sqlx::query_as(
"SELECT f.result, c.result FROM forecasts f JOIN charts c ON c.id = f.chart_id WHERE f.id = $1",
)
.bind(forecast_id)
.fetch_optional(&self.db)
.await
.map_err(|e| LlmError::Transient(format!("database: {e}")))?
.ok_or_else(|| LlmError::Permanent("the forecast was deleted".into()))?;
Ok(compact::forecast(&chart, &result))
}
/// Takes the oldest waiting prompt test run, if any.
pub async fn claim_preview(&self) -> Result<Option<PreviewJob>, sqlx::Error> {
let job: Option<PreviewJob> = sqlx::query_as(
"UPDATE prompt_previews SET status = 'running', started_at = now() \
WHERE id = (SELECT id FROM prompt_previews WHERE status = 'pending' \
ORDER BY created_at LIMIT 1 FOR UPDATE SKIP LOCKED) \
RETURNING id, kind, chart_id, synastry_id, forecast_id, language, model, system",
)
.fetch_optional(&self.db)
.await?;
if let Some(job) = &job {
self.running.lock().unwrap().insert(job.id);
}
Ok(job)
}
/// Writes one test run with the edited system prompt (a single attempt).
pub async fn handle_preview(&self, job: PreviewJob) {
let started = Instant::now();
let result = async {
let kind =
Kind::parse(&job.kind).ok_or_else(|| LlmError::Permanent(format!("unknown kind '{}'", job.kind)))?;
let text = match kind {
Kind::Synastry => self.synastry_text(job.synastry_id).await?,
Kind::Forecast => self.forecast_text(job.forecast_id).await?,
Kind::Natal | Kind::Summary => self.natal_text(job.chart_id).await?,
Kind::Question => return Err(LlmError::Permanent("a question has no test run".into())),
};
let built = prompt::build(kind, &job.language, &text, Some(&job.system));
let mut c = self.llm.complete(&job.model, &built).await?;
prompt::polish(&job.language, &mut c.content);
Ok::<_, LlmError>(c)
}
.await;
let saved = match &result {
Ok(c) => {
sqlx::query(
"UPDATE prompt_previews SET status = 'done', content = $2, model_used = $3, input_tokens = $4, \
output_tokens = $5, cost_usd = $6, completed_at = now() WHERE id = $1 AND status = 'running'",
)
.bind(job.id)
.bind(&c.content)
.bind(&c.model)
.bind(c.input_tokens)
.bind(c.output_tokens)
.bind(c.cost_usd)
.execute(&self.db)
.await
}
Err(e) => {
self.log_error("preview", job.id, &job.kind, &job.model, 1, "failed", e.message(), started.elapsed())
.await;
sqlx::query(
"UPDATE prompt_previews SET status = 'failed', error = $2, completed_at = now() \
WHERE id = $1 AND status = 'running'",
)
.bind(job.id)
.bind(e.message())
.execute(&self.db)
.await
}
};
match (&result, saved) {
(_, Err(e)) => tracing::error!(id = %job.id, error = %e, "saving the prompt preview failed"),
(Ok(c), Ok(_)) => tracing::info!(
id = %job.id, model = %c.model, cost_usd = c.cost_usd, secs = started.elapsed().as_secs(), "prompt preview done"
),
(Err(e), Ok(_)) => tracing::warn!(id = %job.id, error = %e.message(), "prompt preview failed"),
}
self.running.lock().unwrap().remove(&job.id);
}
/// Writes one claimed analysis and records the outcome.
pub async fn handle(&self, job: Job) {
let started = Instant::now();
let work = async {
match self.prompt_for(&job).await {
Ok(prompt) => self.llm.complete(job.active_model(), &prompt).await.map(|mut c| {
prompt::polish(&job.language, &mut c.content);
c
}),
Err(e) => Err(e),
}
};
let result = tokio::select! {
result = work => result,
() = self.superseded("analyses", job.id, job.started_at) => {
tracing::info!(id = %job.id, "analysis restarted by an admin; this attempt is dropped");
return;
}
};
if let Err(e) = self.finish(&job, result, started.elapsed()).await {
tracing::error!(id = %job.id, error = %e, "saving the analysis result failed");
}
self.running.lock().unwrap().remove(&job.id);
}
async fn finish(&self, job: &Job, result: Result<Completion, LlmError>, took: Duration) -> Result<(), sqlx::Error> {
match result {
Ok(c) => {
sqlx::query(
"UPDATE analyses SET status = 'done', content = $2, model_used = $3, input_tokens = $4, \
output_tokens = $5, cost_usd = $6, error = NULL, completed_at = now() \
WHERE id = $1 AND status = 'running' AND started_at = $7",
)
.bind(job.id)
.bind(&c.content)
.bind(&c.model)
.bind(c.input_tokens)
.bind(c.output_tokens)
.bind(c.cost_usd)
.bind(job.started_at)
.execute(&self.db)
.await?;
tracing::info!(
id = %job.id, model = %c.model, input_tokens = c.input_tokens, output_tokens = c.output_tokens,
cost_usd = c.cost_usd, secs = took.as_secs(), "analysis done"
);
}
Err(e) => {
let can_fall_back = !job.on_fallback && job.fallback_model.is_some();
let next = next_step(&e, job.attempts, self.max_attempts, can_fall_back);
let outcome = match next {
Next::Retry(_) => "retry",
Next::Fallback => "fallback",
Next::Fail => "failed",
};
self.log_error(
"analysis",
job.id,
&job.kind,
job.active_model(),
job.attempts,
outcome,
e.message(),
took,
)
.await;
match next {
Next::Retry(wait) => {
sqlx::query(
"UPDATE analyses SET status = 'pending', error = $2, \
next_attempt_at = now() + make_interval(secs => $3) \
WHERE id = $1 AND status = 'running' AND started_at = $4",
)
.bind(job.id)
.bind(e.message())
.bind(wait.as_secs_f64())
.bind(job.started_at)
.execute(&self.db)
.await?;
tracing::warn!(id = %job.id, attempts = job.attempts, retry_in_secs = wait.as_secs(), error = %e.message(), "analysis will be retried");
}
Next::Fallback => {
sqlx::query(
"UPDATE analyses SET status = 'pending', error = $2, on_fallback = true, attempts = 0, \
next_attempt_at = now() WHERE id = $1 AND status = 'running' AND started_at = $3",
)
.bind(job.id)
.bind(e.message())
.bind(job.started_at)
.execute(&self.db)
.await?;
tracing::warn!(
id = %job.id, model = %job.model, fallback = ?job.fallback_model, error = %e.message(),
"the chosen model failed, the default model takes over"
);
}
Next::Fail => {
self.fail(job.id, job.started_at, e.message()).await?;
tracing::error!(id = %job.id, attempts = job.attempts, error = %e.message(), "analysis failed");
}
}
}
}
Ok(())
}
/// Marks the analysis failed and gives its unit back (once).
async fn fail(&self, id: Uuid, started_at: time::OffsetDateTime, message: &str) -> Result<(), sqlx::Error> {
let mut tx = self.db.begin().await?;
let row: Option<(Option<Uuid>, i32, String)> = sqlx::query_as(
"UPDATE analyses SET status = 'failed', error = $2, completed_at = now() \
WHERE id = $1 AND status = 'running' AND started_at = $3 RETURNING requested_by, credits_charged, kind",
)
.bind(id)
.bind(message)
.bind(started_at)
.fetch_optional(&mut *tx)
.await?;
if let Some((Some(user), units, kind)) = row
&& units > 0
{
let unit = match kind.as_str() {
"synastry" => "synastry",
"forecast" => "forecast",
_ => "natal",
};
sqlx::query(
"INSERT INTO credit_ledger (id, user_id, unit, delta, reason, analysis_id, note) \
VALUES ($1, $2, $3, $4, 'refund', $5, 'analysis failed') ON CONFLICT DO NOTHING",
)
.bind(Uuid::now_v7())
.bind(user)
.bind(unit)
.bind(units)
.bind(id)
.execute(&mut *tx)
.await?;
}
tx.commit().await
}
/// Puts analyses left `running` by a stopped worker back in the queue.
pub async fn requeue_stuck(&self) -> Result<u64, sqlx::Error> {
let result = sqlx::query(
"UPDATE analyses SET status = 'pending', error = 'interrupted', next_attempt_at = now() \
WHERE status = 'running' AND started_at < now() - make_interval(mins => $1)",
)
.bind(STUCK_AFTER_MINUTES)
.execute(&self.db)
.await?;
sqlx::query(
"UPDATE questions SET status = 'pending', error = 'interrupted', next_attempt_at = now() \
WHERE status = 'running' AND started_at < now() - make_interval(mins => $1)",
)
.bind(STUCK_AFTER_MINUTES)
.execute(&self.db)
.await?;
// A test run is not retried: the admin can start it again.
sqlx::query(
"UPDATE prompt_previews SET status = 'failed', error = 'interrupted', completed_at = now() \
WHERE status = 'running' AND started_at < now() - make_interval(mins => $1)",
)
.bind(STUCK_AFTER_MINUTES)
.execute(&self.db)
.await?;
Ok(result.rows_affected())
}
/// Records a failed attempt in the admin error log (a failure here is only logged).
#[allow(clippy::too_many_arguments)]
pub(crate) async fn log_error(
&self,
source: &str,
item_id: Uuid,
kind: &str,
model: &str,
attempt: i32,
outcome: &str,
error: &str,
took: Duration,
) {
let result = sqlx::query(
"INSERT INTO generation_errors (source, item_id, kind, model, attempt, outcome, error, seconds) \
VALUES ($1, $2, $3, $4, $5, $6, $7, $8)",
)
.bind(source)
.bind(item_id)
.bind(kind)
.bind(model)
.bind(attempt)
.bind(outcome)
.bind(error.chars().take(2000).collect::<String>())
.bind(i32::try_from(took.as_secs()).unwrap_or(i32::MAX))
.execute(&self.db)
.await;
if let Err(e) = result {
tracing::error!(error = %e, "writing the error log");
}
}
/// Resolves when this attempt is no longer the current one (an admin
/// restarted it); checked every 10 seconds, never on a database error.
async fn superseded(&self, table: &str, id: Uuid, started_at: time::OffsetDateTime) {
let query = format!("SELECT status = 'running' AND started_at = $2 FROM {table} WHERE id = $1");
loop {
tokio::time::sleep(Duration::from_secs(10)).await;
match sqlx::query_scalar::<_, Option<bool>>(&query).bind(id).bind(started_at).fetch_optional(&self.db).await
{
Ok(Some(Some(true))) => {}
Ok(_) => return,
Err(_) => {}
}
}
}
/// On shutdown: analyses still being written go back to `pending` at once.
async fn release_running(&self) {
let ids: Vec<Uuid> = self.running.lock().unwrap().drain().collect();
if ids.is_empty() {
return;
}
let result = sqlx::query(
"UPDATE analyses SET status = 'pending', error = 'interrupted', next_attempt_at = now(), \
attempts = greatest(attempts - 1, 0) WHERE id = ANY($1) AND status = 'running'",
)
.bind(&ids)
.execute(&self.db)
.await;
match result {
Ok(r) => tracing::info!(released = r.rows_affected(), "unfinished analyses put back in the queue"),
Err(e) => tracing::error!(error = %e, "releasing unfinished analyses"),
}
let result = sqlx::query(
"UPDATE questions SET status = 'pending', error = 'interrupted', next_attempt_at = now(), \
attempts = greatest(attempts - 1, 0) WHERE id = ANY($1) AND status = 'running'",
)
.bind(&ids)
.execute(&self.db)
.await;
if let Err(e) = result {
tracing::error!(error = %e, "releasing unfinished questions");
}
let result = sqlx::query(
"UPDATE prompt_previews SET status = 'failed', error = 'interrupted', completed_at = now() \
WHERE id = ANY($1) AND status = 'running'",
)
.bind(&ids)
.execute(&self.db)
.await;
if let Err(e) = result {
tracing::error!(error = %e, "releasing unfinished prompt previews");
}
}
}
/// Work a free slot can take, in order of preference.
pub(crate) enum Work {
Analysis(Job),
Question(QuestionJob),
Preview(PreviewJob),
}
impl Analyzer {
/// Quick work: a question first (someone is waiting for a short answer),
/// then a summary.
pub(crate) async fn claim_quick(&self) -> Result<Option<Work>, sqlx::Error> {
if let Some(job) = self.claim_question().await? {
return Ok(Some(Work::Question(job)));
}
Ok(self.claim(Lane::Quick).await?.map(Work::Analysis))
}
/// Slow work: a detailed analysis, else quick work, else a test run.
pub(crate) async fn claim_slow(&self) -> Result<Option<Work>, sqlx::Error> {
if let Some(job) = self.claim(Lane::Slow).await? {
return Ok(Some(Work::Analysis(job)));
}
if let Some(work) = self.claim_quick().await? {
return Ok(Some(work));
}
Ok(self.claim_preview().await?.map(Work::Preview))
}
async fn work(&self, work: Work) {
match work {
Work::Analysis(job) => self.handle(job).await,
Work::Question(job) => self.handle_question(job).await,
Work::Preview(job) => self.handle_preview(job).await,
}
}
}
pub async fn run(analyzer: Arc<Analyzer>, poll: Duration, stop: watch::Receiver<bool>) {
let quick = Arc::new(Semaphore::new(analyzer.lanes.quick));
let slow = Arc::new(Semaphore::new(analyzer.lanes.slow));
let mut tasks = tokio::task::JoinSet::new();
let mut listener = crate::listen(&analyzer.db, "analyses").await;
let mut last_requeue: Option<Instant> = None;
loop {
if *stop.borrow() {
break;
}
if last_requeue.is_none_or(|t| t.elapsed() > Duration::from_secs(60)) {
last_requeue = Some(Instant::now());
match analyzer.requeue_stuck().await {
Ok(0) => {}
Ok(n) => tracing::warn!(requeued = n, "interrupted analyses requeued"),
Err(e) => tracing::error!(error = %e, "requeueing analyses"),
}
}
// Fill the free slots: the quick lane first, then the slow one.
for lane in [Lane::Quick, Lane::Slow] {
let slots = if lane == Lane::Quick { &quick } else { &slow };
while let Ok(permit) = slots.clone().try_acquire_owned() {
let claimed = match lane {
Lane::Quick => analyzer.claim_quick().await,
Lane::Slow => analyzer.claim_slow().await,
};
match claimed {
Ok(Some(work)) => {
let analyzer = analyzer.clone();
tasks.spawn(async move {
analyzer.work(work).await;
drop(permit);
});
}
Ok(None) => break,
Err(e) => {
tracing::error!(error = %e, ?lane, "claiming work");
break;
}
}
}
}
while tasks.try_join_next().is_some() {}
tokio::select! {
() = crate::notified(listener.as_mut()) => {}
() = tokio::time::sleep(poll) => {}
Some(_) = tasks.join_next(), if !tasks.is_empty() => {}
() = crate::stopped(stop.clone()) => break,
}
}
// Give running analyses a moment, then hand the rest back to the queue.
let _ = tokio::time::timeout(Duration::from_secs(10), async { while tasks.join_next().await.is_some() {} }).await;
tasks.abort_all();
analyzer.release_running().await;
}