Make startup of services more-robust in the face of PostgreSQL not be…

…ing ready. #1393

pscheduler-server/pscheduler-server/daemons/Makefile

@@ -21,6 +21,10 @@ DAEMONS=\
 	ticker \
 	scheduler \
 
+DAEMON_UTILITIES=\
+	wait-for-database
+
+DAEMON_INSTALL=$(DAEMONS) $(DAEMON_UTILITIES)
 
 COMMANDS=\
 	debug \
@@ -123,20 +127,9 @@ endif
 	fi
 TO_CLEAN += resume
 
-
-
-# Systemd 240 and later supports the 'exec' service type, which is
-# preferred becuase it catches more failures sooner.  For anything
-# earlier, fall back on 'simple'.
 SYSTEMD_VERSION := $(shell systemctl --version | awk '$$1 == "systemd" { print $$2 }')
-ifeq ($(shell expr '$(SYSTEMD_VERSION)' '>=' 240), 1)
-     SERVICE_TYPE := exec
-else
-    SERVICE_TYPE := simple
-endif
 
-
-$(UNITS): service-template.raw
+$(UNITS): service-template.m4
 ifndef CONFIGDIR
 	@echo No CONFIGDIR specified for build
 	@false
@@ -169,17 +162,17 @@ ifndef VAR
 	@echo No VAR specified for build
 	@false
 endif
-	sed \
-		-e 's|__CONFIGDIR__|$(CONFIGDIR)|g' \
-		-e 's|__DAEMONDIR__|$(DAEMONDIR)|g' \
-		-e 's|__DSN__|$(DSNFILE)|g' \
-		-e 's|__PROG__|$(@:service-%=%)|g' \
-		-e 's|__PSUSER__|$(PSUSER)|g' \
-		-e 's|__RUNDIR__|$(RUNDIR)|g' \
-		-e 's|__VAR__|$(VAR)|g' \
-		-e 's|__LOGDIR__|$(LOGDIR)|g' \
-		-e 's|__PGSERVICE__|$(PGSERVICE)|g' \
-		-e 's|__SERVICE_TYPE__|$(SERVICE_TYPE)|g' \
+	m4 \
+		-D '__CONFIGDIR__=$(CONFIGDIR)' \
+		-D '__DAEMONDIR__=$(DAEMONDIR)' \
+		-D '__DSN__=$(DSNFILE)' \
+		-D '__PROG__=$(@:service-%=%)' \
+		-D '__PSUSER__=$(PSUSER)' \
+		-D '__RUNDIR__=$(RUNDIR)' \
+		-D '__VAR__=$(VAR)' \
+		-D '__LOGDIR__=$(LOGDIR)' \
+		-D '__PGSERVICE__=$(PGSERVICE)' \
+		-D '__SYSTEMD_VERSION__=$(SYSTEMD_VERSION)' \
 		< $< > $@
 	@if egrep -e '__[A-Z_]+__' $@ ; then \
 		echo "Found un-substituted values in processed file $@" ; \
@@ -199,7 +192,7 @@ $(CONFIGS): config-template.raw
 TO_CLEAN += $(CONFIGS)
 
 
-build: $(DAEMONS) $(UNITS) $(CONFIGS) $(COMMANDS) $(INTERNALS)
+build: $(DAEMON_INSTALL) $(UNITS) $(CONFIGS) $(COMMANDS) $(INTERNALS)
 	@true
 
 
@@ -235,8 +228,8 @@ endif
 	cp -f $(COMMANDS) $(COMMANDDIR)
 	chmod 555 $(COMMANDS:%=$(COMMANDDIR)/%)
 	mkdir -p $(DAEMONDIR)
-	cp -f $(DAEMONS) $(DAEMONDIR)
-	chmod 555 $(DAEMONS:%=$(DAEMONDIR)/%)
+	cp -f $(DAEMON_INSTALL) $(DAEMONDIR)
+	chmod 555 $(DAEMON_INSTALL:%=$(DAEMONDIR)/%)
 	mkdir -p $(INTERNALSDIR)
 	cp -f $(INTERNALS) $(INTERNALSDIR)
 	chmod 555 $(INTERNALS:%=$(INTERNALSDIR)/%)

pscheduler-server/pscheduler-server/daemons/service-template.m4

@@ -0,0 +1,77 @@
+changequote(<<<,>>>)dnl
+changecom()dnl
+#
+# Systemd unit for __PROG__
+#
+# Systemd version __SYSTEMD_VERSION__ was installed at build time.
+#
+# 
+#   Version of systemd installed by distros we support:
+# 
+#     EL7 229
+#     EL8 239
+#     EL9 252
+#     D10 241
+#     D11 247
+#     D12 252
+#     U20 245
+#     U22 249
+
+[Unit]
+Description=pScheduler Server - __PROG__
+# This forces starting and stopping in concert
+PartOf=__PGSERVICE__
+After=__PGSERVICE__
+Wants=__PGSERVICE__
+
+[Service]
+# Systemd 240 added exec, which is better.
+Type=ifelse(eval(__SYSTEMD_VERSION__ <  240),1,simple,exec)
+
+User=__PSUSER__
+Group=__PSUSER__
+
+PermissionsStartOnly=true
+LimitNOFILE=32768
+LimitNPROC=32768
+
+Restart=always
+RestartSec=15
+
+# This is slightly longer than the database check below so failures
+# will be more apparent than just a timeout.
+TimeoutStartSec=130
+
+# Wait for the database to become accessible.  This is done because
+# the PostgreSQL service can appear up when it isn't ready to take
+# queries yet.  That will cause this service to die.
+ExecStartPre=__DAEMONDIR__/wait-for-database --dsn @__DSN__ --dwell 120 --retry 5
+
+# Create the run directory
+ExecStartPre=/bin/mkdir -p __RUNDIR__/__PROG__
+ExecStartPre=/bin/chmod 755 __RUNDIR__/__PROG__
+
+# Set up some temporary space and export its location
+ExecStartPre=/bin/mkdir -p __RUNDIR__/__PROG__/tmp
+ExecStartPre=/bin/chmod 700 __RUNDIR__/__PROG__/tmp
+Environment=TMPDIR=__RUNDIR__/__PROG__/tmp
+
+# Set ownership
+ExecStartPre=/bin/chown -R __PSUSER__:__PSUSER__ __RUNDIR__/__PROG__
+
+# Generate options file
+ExecStartPre=/bin/sh -c "if [ -r __CONFIGDIR__/__PROG__.conf ]; then opts=$(sed -e 's/#.*$//' __CONFIGDIR__/__PROG__.conf); echo OPTIONS=$opts > __RUNDIR__/__PROG__/options; chown __PSUSER__:__PSUSER__ __RUNDIR__/__PROG__/options; fi"
+
+# Redirections
+StandardOutput=journal
+StandardError=journal
+
+# Start service
+EnvironmentFile=-__RUNDIR__/__PROG__/options
+ExecStart=__DAEMONDIR__/__PROG__ --dsn @__DSN__ $OPTIONS
+
+# Stop service
+ExecStopPost=/bin/rm -rf __RUNDIR__/__PROG__
+
+[Install]
+WantedBy=multi-user.target

pscheduler-server/pscheduler-server/daemons/wait-for-database

@@ -0,0 +1,86 @@
+#!/usr/bin/env python3
+#
+# Check that the database is available for connections and working.
+#
+
+import datetime
+import optparse
+import pscheduler
+import psycopg2
+import sys
+import time
+
+pscheduler.set_graceful_exit()
+
+#
+# Gargle the arguments
+#
+
+opt_parser = optparse.OptionParser()
+
+# Program options
+
+opt_parser.add_option("--dsn",
+                      help="Database connection string, prefix with @ to read from file",
+                      action="store", type="string", dest="dsn",
+                      default="")
+opt_parser.add_option("--dwell",
+                      help="How long to keep trying to connect (seconds)",
+                      action="store", type="int", dest="dwell",
+                      default=60)
+opt_parser.add_option("--retry",
+                      help="How often to try connecting (seconds)",
+                      action="store", type="int", dest="retry",
+                      default=2)
+
+(options, args) = opt_parser.parse_args()
+
+
+def try_database(dsn):
+    """
+    Attempt to connect to the database and use it, returning a tuple
+    of True/False if successful/failed and an error message.
+    """
+    db = None
+    try:
+        db = pscheduler.PgConnection(dsn, name="check-database")
+        rows = list(db.query('SELECT 12345'))
+    except Exception as ex:
+        return (False, str(ex))
+    finally:
+        if db is not None:
+            db.close()
+
+    if (len(rows) != 1) or (rows[0] != (12345,)):
+        return(False, 'Got unexpected data: %s' % (rows))
+
+    return (True, 'OK')
+
+
+attempts = 0
+dwell_until = pscheduler.time_now() + datetime.timedelta(seconds=options.dwell)
+
+# This will be overwritten in the loop below.
+reason = 'Never tried to connect'
+
+while pscheduler.time_now() < dwell_until:
+
+    status, reason = try_database(options.dsn)
+
+    if status:
+        print('Successfully connected to the database.', file=sys.stderr)
+        exit(0)
+
+    attempts += 1
+    if attempts == 1:
+        print('Failed first attempt connecting to the database:', reason, file=sys.stderr)
+
+    time.sleep(options.retry)
+
+
+# No dice.
+
+print('Unable to connect within the dwell time.  Last error:\n',
+      reason, file=sys.stderr)
+
+exit(1)

pscheduler-server/pscheduler-server/unibuild-packaging/deb/control

@@ -13,7 +13,7 @@ Architecture: all
 Depends: ${misc:Depends}, python3, curl,
  python3-daemon, python3-flask, python3-psutil,
  python3-tz, python3-jsontemplate, python3-radix, python3-crontab,
- python3-pscheduler (>= 4.4.0~), pscheduler-core, pscheduler-account,
+ python3-pscheduler (>= 5.0.8~), pscheduler-core, pscheduler-account,
  apache2, libapache2-mod-wsgi-py3, logrotate, psmisc,
  dbconfig-common, postgresql (>= 9.5),
  postgresql-12 | postgresql-11 | postgresql-10 | postgresql-9.6 | postgresql-9.5,

pscheduler-server/pscheduler-server/unibuild-packaging/rpm/pscheduler-server.spec

@@ -72,7 +72,7 @@ Requires:	httpd-wsgi-socket
 Requires:	mod_ssl
 Requires:	mod_wsgi > 4.0
 Requires:	%{_pscheduler_python}-parse-crontab
-Requires:	%{_pscheduler_python}-pscheduler >= 5.0.0
+Requires:	%{_pscheduler_python}-pscheduler >= 5.0.8
 
 %if 0%{?el7}
 Requires:	pytz

python-pscheduler/pscheduler/pscheduler/db.py

@@ -96,6 +96,9 @@ def __init__(self, dsn, autocommit=True, name=None):
         self.pending_notifications = {}
         self.lock = threading.Lock()
 
+    def close(self):
+        self.pg.close()
+
     #
     # Notifications
     #