From 8955727738a9aff96c47e4b13d7713eacd91c3ce Mon Sep 17 00:00:00 2001
From: Danny Cocks <danny.cocks@canonical.com>
Date: Mon, 23 Oct 2023 16:06:31 +1100
Subject: [PATCH] Switch the deletion of the cron output with a check for a
 time

The removal of the cron job output messes up the behaviour of the NRPE check if
it called more than once in a minute. The intention of removing the cron job
output is to prevent stale data from a false STATE_OK. I've rewritten it so the
cron job puts the time into the file so stale output can be detected.

I went with explicit output in the file as I've known of issues with mtime
comparisons in the past.

Closes-Bug: #2019251
Change-Id: I91d0430eecb58b6e1a1d1ef1d75e38d88ebfd30d
---
 files/nagios/check_ceph_osd_services.py   | 28 ++++++++++++++++++-----
 files/nagios/collect_ceph_osd_services.py |  5 ++++
 2 files changed, 27 insertions(+), 6 deletions(-)

diff --git a/files/nagios/check_ceph_osd_services.py b/files/nagios/check_ceph_osd_services.py
index 7f53b2d7..28669c28 100755
--- a/files/nagios/check_ceph_osd_services.py
+++ b/files/nagios/check_ceph_osd_services.py
@@ -6,9 +6,13 @@
 
 import os
 import sys
+import time
 
 CRON_CHECK_TMPFILE = 'ceph-osd-checks'
 NAGIOS_HOME = '/var/lib/nagios'
+CRON_CHECK_PREFIX = 'Ceph OSD cron check:'
+# The cron job runs every minute, so allow for 2 min
+CRON_TIMEOUT = 120
 
 STATE_OK = 0
 STATE_WARNING = 1
@@ -38,16 +42,28 @@ def run_main():
         print("Something went wrong reading the file: {}".format(str(e)))
         return STATE_UNKNOWN
 
-    # now remove the file in case the next check fails.
-    try:
-        os.remove(_tmp_file)
-    except Exception:
-        pass
-
     if not lines:
         print("checked status file is empty: {}".format(_tmp_file))
         return STATE_UNKNOWN
 
+    # check the first line to see if is the expected time the output is stale
+    firstline = lines.pop(0)
+    if not firstline.startswith(CRON_CHECK_PREFIX):
+        print("First line not of the expected format: {}".format(firstline))
+        return STATE_UNKNOWN
+
+    try:
+        rest_of_line = firstline[len(CRON_CHECK_PREFIX):]
+        cron_time = int(rest_of_line)
+    except Exception as exc:
+        print("Unable to parse time from first line: {}".format(exc))
+        return STATE_UNKNOWN
+
+    if cron_time + CRON_TIMEOUT < time.time():
+        how_old = int(time.time() - cron_time)
+        print("Cron output is stale ({} sec old)".format(how_old))
+        return STATE_UNKNOWN
+
     # finally, check that the file contains all ok lines.  Unfortunately, it's
     # not consistent across releases, but what is consistent is that the check
     # command in the collect phase does fail, and so the start of the line is
diff --git a/files/nagios/collect_ceph_osd_services.py b/files/nagios/collect_ceph_osd_services.py
index 633148a2..fae8e774 100755
--- a/files/nagios/collect_ceph_osd_services.py
+++ b/files/nagios/collect_ceph_osd_services.py
@@ -6,6 +6,7 @@
 
 import os
 import subprocess
+import time
 from pwd import getpwnam
 
 # fasteners only exists in Bionic, so this will fail on xenial and trusty
@@ -18,6 +19,7 @@ SYSTEMD_SYSTEM = '/run/systemd/system'
 LOCKFILE = '/var/lock/check-osds.lock'
 CRON_CHECK_TMPFILE = 'ceph-osd-checks'
 NAGIOS_HOME = '/var/lib/nagios'
+CRON_CHECK_PREFIX = 'Ceph OSD cron check:'
 
 
 def init_is_systemd():
@@ -57,6 +59,9 @@ def do_status():
 
     lines = []
 
+    # First line contains the time so the checker can identify stale data
+    lines.append("{} {}\n".format(CRON_CHECK_PREFIX, int(time.time())))
+
     for unit in get_osd_units():
         try:
             output = (subprocess