diff --git a/include/fluent-bit/flb_log.h b/include/fluent-bit/flb_log.h
index 75de5060984..90f145efff8 100644
--- a/include/fluent-bit/flb_log.h
+++ b/include/fluent-bit/flb_log.h
@@ -35,7 +35,11 @@
 #include <stdarg.h>
 
 /* FIXME: this extern should be auto-populated from flb_thread_storage.h */
-extern FLB_TLS_DEFINE(struct flb_log, flb_log_ctx)
+#ifndef FLB_HAVE_C_TLS
+FLB_TLS_DECLARE(struct flb_log, flb_log_ctx);
+#else
+extern FLB_TLS_DEFINE(struct flb_log, flb_log_ctx);
+#endif
 
 /* Message types */
 #define FLB_LOG_OFF     0
diff --git a/include/fluent-bit/flb_output.h b/include/fluent-bit/flb_output.h
index d9675e48bc4..7cd3a9628f1 100644
--- a/include/fluent-bit/flb_output.h
+++ b/include/fluent-bit/flb_output.h
@@ -592,7 +592,11 @@ struct flb_out_flush_params {
     struct flb_coro *coro;                      /* coroutine context     */
 };
 
+#ifndef FLB_HAVE_C_TLS
+FLB_TLS_DECLARE(struct flb_out_flush_params, out_flush_params);
+#else
 extern FLB_TLS_DEFINE(struct flb_out_flush_params, out_flush_params);
+#endif
 
 #define FLB_OUTPUT_RETURN(x)                                            \
     flb_output_return_do(x);                                            \
diff --git a/include/fluent-bit/flb_scheduler.h b/include/fluent-bit/flb_scheduler.h
index 11b526d350f..f5ecc042310 100644
--- a/include/fluent-bit/flb_scheduler.h
+++ b/include/fluent-bit/flb_scheduler.h
@@ -200,7 +200,11 @@ struct flb_sched_timer_coro_cb_params {
     struct flb_coro *coro;
 };
 
+#ifndef FLB_HAVE_C_TLS
+FLB_TLS_DECLARE(struct flb_sched_timer_coro_cb_params, sched_timer_coro_cb_params);
+#else
 extern FLB_TLS_DEFINE(struct flb_sched_timer_coro_cb_params, sched_timer_coro_cb_params);
+#endif
 
 
 struct flb_timer_cb_coro_params {
diff --git a/include/fluent-bit/flb_thread_storage.h b/include/fluent-bit/flb_thread_storage.h
index 7ec9df94010..a442929f96f 100644
--- a/include/fluent-bit/flb_thread_storage.h
+++ b/include/fluent-bit/flb_thread_storage.h
@@ -40,13 +40,41 @@
 /* Fallback mode using pthread_*() for Thread-Local-Storage usage */
 #define FLB_TLS_SET(key, val)      pthread_setspecific(key, (void *) val)
 #define FLB_TLS_GET(key)           pthread_getspecific(key)
-#define FLB_TLS_INIT(key)          pthread_key_create(&key, NULL)
-#define FLB_TLS_DEFINE(type, name) pthread_key_t name;
+
+/*
+ * Thread-safe idempotent initialization using pthread_once.
+ * This ensures pthread_key_create is only called once even if FLB_TLS_INIT
+ * is called from multiple locations (different compilation units, hot reload, etc).
+ */
+#define FLB_TLS_INIT(key) \
+    do { \
+        extern pthread_once_t key##_once; \
+        void key##_init_func(void); \
+        pthread_once(&key##_once, key##_init_func); \
+    } while(0)
+
+/* Define a TLS key with its pthread_once control and init function */
+#define FLB_TLS_DEFINE(type, name) \
+    pthread_key_t name; \
+    pthread_once_t name##_once = PTHREAD_ONCE_INIT; \
+    void name##_init_func(void) { \
+        pthread_key_create(&name, NULL); \
+    }
+
+/* Declare a TLS key that's defined elsewhere */
+#define FLB_TLS_DECLARE(type, name) \
+    extern pthread_key_t name; \
+    extern pthread_once_t name##_once; \
+    void name##_init_func(void);
 #endif
 
 
 /* FIXME: this extern should be auto-populated from flb_thread_storage.h */
-extern FLB_TLS_DEFINE(struct flb_worker, flb_worker_ctx)
+#ifndef FLB_HAVE_C_TLS
+FLB_TLS_DECLARE(struct flb_worker, flb_worker_ctx);
+#else
+extern FLB_TLS_DEFINE(struct flb_worker, flb_worker_ctx);
+#endif
 
 
 #endif /* !FLB_THREAD_STORAGE_H */