fix(logging): fsync crash logs before _Exit() to prevent data loss

Leiyks · Leiyks · commit acfde3da7a73 · 2026-02-11T13:34:38.000+01:00
When a SIGSEGV occurs, the signal handler logs "Segmentation fault encountered"
and then calls _Exit() which terminates the process immediately. Without fsync(),
kernel write buffers may not be flushed to disk before termination, causing
a race condition where the error log file is sometimes not created.

This fix adds fsync() on Unix/Linux and _commit() on Windows after write() in
ddtrace_log_with_time() to ensure crash logs persist to disk before process
termination.

The issue affects production (rare but possible during power loss, kernel panic,
or I/O errors) and causes consistent test failures where tests check for log
files immediately after crashes (before kernel writeback completes).

Fixes flaky test_metrics SigSegVTest::testGet failures on Kubernetes where
dd_php_error.log was not being created consistently.
diff --git a/ext/logging.c b/ext/logging.c
@@ -160,6 +160,15 @@ int ddtrace_log_with_time(int fd, const char *msg, int msg_len) {
 
     int ret = write(fd, msgbuf, p - msgbuf);
 
+    // Flush to disk to ensure data persists even if process terminates immediately
+    if (ret > 0) {
+#ifndef _WIN32
+        fsync(fd);
+#else
+        _commit(fd);
+#endif
+    }
+
     free(msgbuf);
     return ret;
 }
