diff --git a/postgres/changelog.d/23728.fixed b/postgres/changelog.d/23728.fixed
new file mode 100644
index 0000000000000..aec07d0c61155
--- /dev/null
+++ b/postgres/changelog.d/23728.fixed
@@ -0,0 +1 @@
+Fix a crash caused by cancel closing database connections while the check is still running.
\ No newline at end of file
diff --git a/postgres/datadog_checks/postgres/postgres.py b/postgres/datadog_checks/postgres/postgres.py
index dac424e3f9c5f..dbd4bd067bc7c 100644
--- a/postgres/datadog_checks/postgres/postgres.py
+++ b/postgres/datadog_checks/postgres/postgres.py
@@ -5,6 +5,7 @@
 import copy
 import functools
 import os
+import threading
 from string import Template
 from time import time
 
@@ -17,7 +18,6 @@
 from datadog_checks.base.utils.db.core import QueryManager
 from datadog_checks.base.utils.db.health import HealthEvent, HealthStatus
 from datadog_checks.base.utils.db.utils import (
-    DBMAsyncJob,
     default_json_event_encoding,
     tracked_query,
 )
@@ -194,6 +194,10 @@ def __init__(self, name, init_config, instances):
 
         self.diagnosis.register(functools.partial(run_diagnostics, self))
 
+        self._cancel_lock = threading.Lock()
+        self._is_running = False
+        self._cancelled = False
+
     def database_monitoring_column_statistics(self, raw_event: str):
         self.event_platform_event(raw_event, "dbm-column-statistics")
 
@@ -476,38 +480,87 @@ def dynamic_queries(self):
 
         return self._dynamic_queries
 
-    @staticmethod
-    def _cancel_async_job(job: DBMAsyncJob):
-        job.cancel()
-        if job._job_loop_future:
-            job._job_loop_future.result()
-            job._job_loop_future = None
-        job._shutdown()
+    def run(self):
+        # TODO: move this lock into the base class
+        with self._cancel_lock:
+            if self._cancelled:
+                self.log.debug("run() skipped, check already cancelled")
+                return ''
+            self._is_running = True
+        try:
+            return super().run()
+        finally:
+            needs_finalize = False
+            with self._cancel_lock:
+                self._is_running = False
+                if self._cancelled:
+                    needs_finalize = True
+            if needs_finalize:
+                self.log.debug("Check cancel has been signaled, finalizing now that run() is complete")
+                self._finalize()
 
     def cancel(self):
+        """Signal that the check is being unscheduled.
+
+        This method can be called while check() is running on another thread
+        (the GIL is released during psycopg I/O). It must not perform any
+        destructive operations — closing connections or nulling attributes that
+        check() depends on — because that causes a SIGSEGV in libpq when
+        check() resumes.
+
+        Destructive cleanup is deferred to _finalize(), which is called either
+        here (if the check is idle) or by run()'s finally block (if the check
+        is in-flight). The Agent guarantees it will not call run() again after
+        cancel().
         """
-        Cancels and sends cancel signal to all threads.
-        """
+        self.log.debug("Marking check as cancelled")
+        self._cancel_async_jobs()
+        needs_finalize = False
+        with self._cancel_lock:
+            self._cancelled = True
+            if not self._is_running:
+                needs_finalize = True
+        if needs_finalize:
+            self.log.debug("cancel() finalizing immediately, check is idle")
+            self._finalize()
+        else:
+            self.log.debug("cancel() deferred finalize, check is still running")
+
+    @property
+    def _async_jobs(self):
+        """Return the async jobs active for this check's configuration."""
+        jobs = []
         if self._config.dbm:
-            self._cancel_async_job(self.statement_metrics)
-            self._cancel_async_job(self.statement_samples)
-            self._cancel_async_job(self.metadata_samples)
+            jobs.extend([self.statement_metrics, self.statement_samples, self.metadata_samples])
         elif self._config.data_observability.enabled:
-            self._cancel_async_job(self.metadata_samples)
+            jobs.append(self.metadata_samples)
         if self._config.data_observability.enabled:
-            self._cancel_async_job(self.data_observability)
+            jobs.append(self.data_observability)
+        return jobs
+
+    def _cancel_async_jobs(self):
+        """Signal async jobs to stop. Safe to call while check() is running."""
+        for job in self._async_jobs:
+            job.cancel()
+
+    def _finalize(self):
+        """Tear down check state. Must not run while check() is executing."""
+        self.log.debug("Finalizing check: closing connections and clearing state")
+        for job in self._async_jobs:
+            if job._job_loop_future:
+                job._job_loop_future.result()
+                job._job_loop_future = None
+            job._shutdown()
         self._clean_state()
-        self._query_manager = None
-        self.health = None
         self.check_initializations.clear()
         # TODO: move diagnosis cleanup into AgentCheck.cancel() in the base class
         self._diagnosis = None
+        self.log.check = None
+        self._query_manager = None
+        self.health = None
         self._close_db()
         self._close_db_pool()
-        # CheckLoggingAdapter holds self.check until check_id is resolved via
-        # process(), which only happens after the agent scheduler calls run().
-        # If cancel() is called before that, the back-reference is never cleared.
-        self.log.check = None
+        self.log.debug("Check cleanup complete")
 
     def _clean_state(self):
         self.log.debug("Cleaning state")
@@ -1191,14 +1244,15 @@ def check(self, _):
 
             if not self._config.only_custom_queries:
                 self._collect_stats(tags)
-                if self._config.dbm:
-                    self.statement_metrics.run_job_loop(tags)
-                    self.statement_samples.run_job_loop(tags)
-                    self.metadata_samples.run_job_loop(tags)
-                elif self._config.data_observability.enabled:
-                    self.metadata_samples.run_job_loop(tags)
-                if self._config.data_observability.enabled:
-                    self.data_observability.run_job_loop(tags)
+                if not self._cancelled:
+                    if self._config.dbm:
+                        self.statement_metrics.run_job_loop(tags)
+                        self.statement_samples.run_job_loop(tags)
+                        self.metadata_samples.run_job_loop(tags)
+                    elif self._config.data_observability.enabled:
+                        self.metadata_samples.run_job_loop(tags)
+                    if self._config.data_observability.enabled:
+                        self.data_observability.run_job_loop(tags)
                 if self._config.collect_wal_metrics is True:
                     # collect wal metrics for pg < 10 only when explicitly enabled
                     # (requires local filesystem access to the WAL directory)
diff --git a/postgres/tests/test_unit.py b/postgres/tests/test_unit.py
index 7efdb71479139..5505131e71ce1 100644
--- a/postgres/tests/test_unit.py
+++ b/postgres/tests/test_unit.py
@@ -443,6 +443,80 @@ def test_check_gc_after_cancel(pg_instance):
         gc.enable()
 
 
+def test_cancel_during_running_check_defers_finalize(pg_instance):
+    """Verify that cancel() during an in-flight check() does not close connections.
+
+    Destructive cleanup (_finalize) must be deferred until run() completes so
+    that check() never accesses a closed psycopg connection, which would cause
+    a SIGSEGV in libpq.
+    """
+    import threading
+
+    check = PostgreSql('postgres', {}, [pg_instance])
+    conn = mock.MagicMock()
+    check._db = conn
+
+    check_started = threading.Event()
+    cancel_done = threading.Event()
+
+    def slow_run(self_arg):
+        check_started.set()
+        cancel_done.wait(timeout=5)
+        return ''
+
+    run_result = [None]
+
+    def run_check():
+        with mock.patch.object(type(check).__mro__[1], 'run', slow_run):
+            run_result[0] = check.run()
+
+    run_thread = threading.Thread(target=run_check)
+    run_thread.start()
+
+    check_started.wait(timeout=5)
+
+    check.cancel()
+    # cancel() should have signaled but NOT finalized since run() is in-flight
+    assert not conn.close.called, "_close_db() ran while check() was still executing"
+    assert check._cancelled is True
+
+    cancel_done.set()
+    run_thread.join(timeout=5)
+
+    # After run() completes, _finalize() should have been called
+    conn.close.assert_called_once()
+    assert check._db is None
+    assert check._query_manager is None
+    assert check.health is None
+
+
+def test_cancel_on_idle_check_finalizes_immediately(pg_instance):
+    """Verify that cancel() on an idle check runs _finalize() inline."""
+    check = PostgreSql('postgres', {}, [pg_instance])
+    conn = mock.MagicMock()
+    check._db = conn
+
+    assert not check._is_running
+
+    check.cancel()
+
+    conn.close.assert_called_once()
+    assert check._db is None
+    assert check._query_manager is None
+    assert check.health is None
+
+
+def test_run_after_cancel_returns_immediately(pg_instance):
+    """Verify that run() returns '' without executing check() if already cancelled."""
+    check = PostgreSql('postgres', {}, [pg_instance])
+    check.cancel()
+
+    with mock.patch.object(check, 'check', side_effect=AssertionError("check() should not be called")):
+        result = check.run()
+
+    assert result == ''
+
+
 def test_collect_column_statistics_updates_timestamp_on_failure(pg_instance):
     pg_instance['dbm'] = True
     pg_instance['collect_column_statistics'] = {'enabled': True, 'collection_interval': 60}