feat(metrics): count api requests auth failures and 5xxs

ts/api/server.ts

@@ -31,6 +31,9 @@ export class ApiServer {
   private clusterCoordinator: ClusterCoordinator;
   private clusterHandler: ClusterHandler;
   private startTime: number = 0;
+  private requestCounts = new Map<string, number>();
+  private authFailureCounts = new Map<string, number>();
+  private serverErrorCounts = new Map<string, number>();
 
   constructor(
     config: IApiConfig,
@@ -131,18 +134,21 @@ export class ApiServer {
 
     if (path.startsWith('/_cluster')) {
       await this.clusterHandler.handle(req, res, path, url);
+      this.recordRequest(path, res.statusCode);
       return;
     }
 
     // Health check endpoint (no auth required)
     if (path === '/health' || path === '/healthz') {
       await this.handleHealthCheck(res);
+      this.recordRequest(path, res.statusCode);
       return;
     }
 
     // Metrics endpoint (no auth required)
     if (path === '/metrics') {
       await this.handleMetrics(res);
+      this.recordRequest(path, res.statusCode);
       return;
     }
 
@@ -156,6 +162,7 @@ export class ApiServer {
 
     // Log request
     const duration = Date.now() - startTime;
+    this.recordRequest(path, res.statusCode);
     logger.dim(`${req.method} ${path} - ${res.statusCode} (${duration}ms)`);
   }
 
@@ -268,6 +275,28 @@ export class ApiServer {
       metrics.push(`# TYPE modelgrid_gpus_total gauge`);
       metrics.push(`modelgrid_gpus_total ${gpus.length}`);
 
+      for (const [path, count] of this.requestCounts.entries()) {
+        metrics.push(`# HELP modelgrid_api_requests_total Total API requests by path`);
+        metrics.push(`# TYPE modelgrid_api_requests_total counter`);
+        metrics.push(`modelgrid_api_requests_total{path="${this.escapeMetricLabel(path)}"} ${count}`);
+      }
+
+      for (const [path, count] of this.authFailureCounts.entries()) {
+        metrics.push(`# HELP modelgrid_api_auth_failures_total Total authentication failures by path`);
+        metrics.push(`# TYPE modelgrid_api_auth_failures_total counter`);
+        metrics.push(
+          `modelgrid_api_auth_failures_total{path="${this.escapeMetricLabel(path)}"} ${count}`,
+        );
+      }
+
+      for (const [path, count] of this.serverErrorCounts.entries()) {
+        metrics.push(`# HELP modelgrid_api_server_errors_total Total 5xx responses by path`);
+        metrics.push(`# TYPE modelgrid_api_server_errors_total counter`);
+        metrics.push(
+          `modelgrid_api_server_errors_total{path="${this.escapeMetricLabel(path)}"} ${count}`,
+        );
+      }
+
       res.writeHead(200, { 'Content-Type': 'text/plain; charset=utf-8' });
       res.end(metrics.join('\n') + '\n');
     } catch (error) {
@@ -310,4 +339,24 @@ export class ApiServer {
       uptime: this.startTime ? Math.floor((Date.now() - this.startTime) / 1000) : 0,
     };
   }
+
+  private recordRequest(path: string, statusCode: number): void {
+    this.incrementMetric(this.requestCounts, path);
+
+    if (statusCode === 401) {
+      this.incrementMetric(this.authFailureCounts, path);
+    }
+
+    if (statusCode >= 500) {
+      this.incrementMetric(this.serverErrorCounts, path);
+    }
+  }
+
+  private incrementMetric(metric: Map<string, number>, path: string): void {
+    metric.set(path, (metric.get(path) || 0) + 1);
+  }
+
+  private escapeMetricLabel(value: string): string {
+    return value.replaceAll('\\', '\\\\').replaceAll('"', '\\"');
+  }
 }