наблюдаемость: завершить базовый контур Community
CI / Rust Checks (push) Failing after 4m28s
CI / UI Checks (push) Has been skipped
CI / Frontend E2E (push) Has been skipped
CI / Community Image Smoke (push) Has been skipped
CI / Deploy (push) Has been skipped

Добавить структурированные журналы, метрики, трассировку и безопасный канал критических ошибок. Усилить границы рантайма, тесты, проверку зависимостей и сценарии развёртывания.
This commit is contained in:
2026-07-31 01:01:14 +03:00
parent 99bd05c145
commit 0e8f1ca03a
160 changed files with 13506 additions and 1499 deletions
+137 -47
View File
@@ -1,24 +1,27 @@
use std::{
collections::HashMap,
sync::Arc,
sync::{Arc, Weak},
time::{Duration, Instant, SystemTime, UNIX_EPOCH},
};
use crank_core::{CacheScope, CoordinationStateStore, CoordinationStateValue};
use crank_registry::{PostgresRegistry, PublishedAgentCatalog, PublishedAgentTool, RegistryError};
use crank_trace::{DbOperation, ErrorCategory, Stage, StageOutcome};
use serde::{Deserialize, Serialize};
use tokio::sync::{Mutex, RwLock};
use tracing::{info, warn};
use tracing::{Instrument, info, warn};
use crate::manifest::analyze_published_tool_catalog;
const MAX_LOCAL_CATALOGS: usize = 1_024;
#[derive(Clone)]
pub struct PublishedToolCatalog {
registry: PostgresRegistry,
refresh_interval: Duration,
coordination_store: Arc<dyn CoordinationStateStore>,
cached: Arc<RwLock<HashMap<CatalogKey, CachedCatalog>>>,
refresh_locks: Arc<Mutex<HashMap<CatalogKey, Arc<Mutex<()>>>>>,
refresh_locks: Arc<Mutex<HashMap<CatalogKey, Weak<Mutex<()>>>>>,
}
#[derive(Clone, Debug, PartialEq, Eq, Hash)]
@@ -30,6 +33,14 @@ struct CatalogKey {
struct CachedCatalog {
loaded_at: Option<Instant>,
catalog: PublishedAgentCatalog,
metrics: CatalogMetrics,
}
#[derive(Clone, Copy, Debug, Default)]
struct CatalogMetrics {
tool_count: usize,
estimated_context_tokens: usize,
warning_count: usize,
}
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
@@ -66,15 +77,29 @@ impl PublishedToolCatalog {
workspace_slug: &str,
agent_slug: &str,
) -> Result<PublishedAgentCatalog, RegistryError> {
self.refresh_if_stale(workspace_slug, agent_slug).await?;
let guard = self.cached.read().await;
guard
.get(&CatalogKey::new(workspace_slug, agent_slug))
.map(|entry| entry.catalog.clone())
.ok_or_else(|| RegistryError::PublishedAgentNotFound {
workspace_slug: workspace_slug.to_owned(),
agent_slug: agent_slug.to_owned(),
})
let span = Stage::McpCatalogLoad.span();
let result = async {
self.refresh_if_stale(workspace_slug, agent_slug).await?;
let guard = self.cached.read().await;
guard
.get(&CatalogKey::new(workspace_slug, agent_slug))
.map(|entry| entry.catalog.clone())
.ok_or_else(|| RegistryError::PublishedAgentNotFound {
workspace_slug: workspace_slug.to_owned(),
agent_slug: agent_slug.to_owned(),
})
}
.instrument(span.clone())
.await;
match &result {
Ok(_) => StageOutcome::Success.record(&span),
Err(_) => {
StageOutcome::Error.record(&span);
ErrorCategory::Catalog.record(&span);
}
}
drop(span);
result
}
async fn refresh_if_stale(
@@ -98,11 +123,14 @@ impl PublishedToolCatalog {
let refresh_lock = {
let mut locks = self.refresh_locks.lock().await;
Arc::clone(
locks
.entry(key.clone())
.or_insert_with(|| Arc::new(Mutex::new(()))),
)
locks.retain(|_, lock| lock.strong_count() > 0);
if let Some(lock) = locks.get(&key).and_then(Weak::upgrade) {
lock
} else {
let lock = Arc::new(Mutex::new(()));
locks.insert(key.clone(), Arc::downgrade(&lock));
lock
}
};
let _refresh_guard = refresh_lock.lock().await;
let still_stale = {
@@ -117,50 +145,59 @@ impl PublishedToolCatalog {
}
if let Some((catalog, age)) = self.load_shared_snapshot(workspace_slug, agent_slug).await {
log_catalog_analysis(workspace_slug, agent_slug, "shared_cache", &catalog.tools);
let mut guard = self.cached.write().await;
guard.insert(
let metrics =
log_catalog_analysis(workspace_slug, agent_slug, "shared_cache", &catalog.tools);
self.store_local_catalog(
key,
CachedCatalog {
loaded_at: Instant::now().checked_sub(age),
catalog,
metrics,
},
);
)
.await;
return Ok(());
}
let catalog = match self
let db_span = Stage::DbQuery
.db_span(DbOperation::CatalogLoad)
.expect("database stage");
let catalog_result = self
.registry
.get_published_agent_catalog_by_slug(workspace_slug, agent_slug)
.await
{
.instrument(db_span.clone())
.await;
let catalog = match catalog_result {
Ok(catalog) => catalog,
Err(error) => return Err(error),
Err(error) => {
StageOutcome::Error.record(&db_span);
ErrorCategory::Database.record(&db_span);
drop(db_span);
return Err(error);
}
};
log_catalog_analysis(workspace_slug, agent_slug, "postgres", &catalog.tools);
StageOutcome::Success.record(&db_span);
drop(db_span);
let metrics = log_catalog_analysis(workspace_slug, agent_slug, "postgres", &catalog.tools);
self.store_shared_snapshot(workspace_slug, agent_slug, &catalog)
.await;
let mut guard = self.cached.write().await;
let previous_count = guard
.get(&key)
.map(|entry| entry.catalog.tools.len())
.unwrap_or_default();
guard.insert(
key,
CachedCatalog {
loaded_at: Some(Instant::now()),
catalog,
},
);
let published_tool_count = catalog.tools.len();
let previous_count = self
.store_local_catalog(
key,
CachedCatalog {
loaded_at: Some(Instant::now()),
catalog,
metrics,
},
)
.await;
info!(
name: "mcp.catalog.refreshed",
workspace_slug,
agent_slug,
published_tool_count = guard
.get(&CatalogKey::new(workspace_slug, agent_slug))
.map(|entry| entry.catalog.tools.len())
.unwrap_or_default(),
published_tool_count,
previous_published_tool_count = previous_count,
"published agent catalog refreshed"
);
@@ -168,6 +205,27 @@ impl PublishedToolCatalog {
Ok(())
}
async fn store_local_catalog(&self, key: CatalogKey, entry: CachedCatalog) -> usize {
let mut guard = self.cached.write().await;
let previous_count = guard
.get(&key)
.map(|current| current.catalog.tools.len())
.unwrap_or_default();
if guard.len() >= MAX_LOCAL_CATALOGS && !guard.contains_key(&key) {
let oldest = guard
.iter()
.min_by_key(|(_, current)| current.loaded_at)
.map(|(candidate, _)| candidate.clone());
if let Some(oldest) = oldest {
guard.remove(&oldest);
}
}
guard.insert(key, entry);
record_catalog_metrics(guard.values().map(|entry| entry.metrics));
previous_count
}
async fn load_shared_snapshot(
&self,
workspace_slug: &str,
@@ -226,12 +284,19 @@ fn log_catalog_analysis(
agent_slug: &str,
source: &str,
tools: &[PublishedAgentTool],
) {
) -> CatalogMetrics {
let analysis = match analyze_published_tool_catalog(tools) {
Ok(analysis) => analysis,
Err(error) => {
warn!(workspace_slug, agent_slug, source, %error, "published catalog analysis failed");
return;
Err(_) => {
warn!(
name: "mcp.catalog.analysis_failed",
workspace_slug,
agent_slug,
source,
error_category = "catalog_validation",
"published catalog analysis failed"
);
return CatalogMetrics::default();
}
};
let warning_count = analysis
@@ -242,6 +307,7 @@ fn log_catalog_analysis(
.count();
info!(
name: "mcp.catalog.analyzed",
workspace_slug,
agent_slug,
source,
@@ -255,6 +321,30 @@ fn log_catalog_analysis(
catalog_quality_warning_count = warning_count,
"published agent catalog analyzed"
);
CatalogMetrics {
tool_count: analysis.budget.tool_count,
estimated_context_tokens: analysis.budget.estimated_context_tokens,
warning_count,
}
}
fn record_catalog_metrics(metrics: impl Iterator<Item = CatalogMetrics>) {
let aggregate = metrics.fold(CatalogMetrics::default(), |mut aggregate, current| {
aggregate.tool_count = aggregate.tool_count.saturating_add(current.tool_count);
aggregate.estimated_context_tokens = aggregate
.estimated_context_tokens
.saturating_add(current.estimated_context_tokens);
aggregate.warning_count = aggregate
.warning_count
.saturating_add(current.warning_count);
aggregate
});
metrics::gauge!("crank_catalog_tools").set(aggregate.tool_count as f64);
metrics::gauge!("crank_catalog_estimated_context_tokens")
.set(aggregate.estimated_context_tokens as f64);
metrics::gauge!("crank_catalog_warnings").set(aggregate.warning_count as f64);
}
fn now_unix_ms() -> u64 {