configurable health threshold

crates/service/job/inference_job/src/http_server/endpoints/health_check_handler.rs

@@ -63,9 +63,6 @@ pub async fn get_health_check_handler(
         HealthCheckError::ServerError
       })?;
 
-  // TODO: Configurability
-  let is_healthy = job_stats.consecutive_failure_count < 10;
-
   let total_tries = job_stats.total_failure_count.saturating_add(job_stats.total_success_count);
 
   let total_success_ratio = if total_tries > 0 {
@@ -80,6 +77,9 @@ pub async fn get_health_check_handler(
     0.0
   };
 
+  let is_healthy =
+      job_stats.consecutive_failure_count < server_state.consecutive_failure_unhealthy_threshold;
+
   let response = HealthCheckResponse {
     success: true,
     is_healthy,

crates/service/job/inference_job/src/http_server/http_server_shared_state.rs

@@ -3,4 +3,5 @@ use jobs_common::job_stats::JobStats;
 #[derive(Clone)]
 pub struct HttpServerSharedState {
   pub job_stats: JobStats,
+  pub consecutive_failure_unhealthy_threshold: u64,
 }

crates/service/job/inference_job/src/http_server/run_http_server.rs

@@ -27,6 +27,8 @@ pub fn run_http_server(args: CreateServerArgs) -> AnyhowResult<Server>
 
   let server_state = HttpServerSharedState {
     job_stats: args.job_stats.clone(),
+    consecutive_failure_unhealthy_threshold:
+      easyenv::get_env_num("CONSECUTIVE_FAILURE_UNHEALTHY_THRESHOLD", 3)?,
   };
 
   let server_state_arc = web::Data::new(Arc::new(server_state));