Merge lp:~openerp-dev/openobject-server/trunk-gunicorn-signaling-vmt into lp:openobject-server

[Antony Lesuisse (OpenERP)]

multiprocess signaling, wip

[Vo Minh Thu (thu)]

We have to add --fixes for this bug report:
https://bugs.launchpad.net/openobject-server/+bug/943794

[lp:~openerp-dev/openobject-server/trunk-gunicorn-signaling-vmt]

4018. By Vo Minh Thu on 2012-03-07

[IMP] multi-process: moved signaling sequences to registry creation instead of base.sql.

4019. By Vo Minh Thu on 2012-03-07

[MERGE] merged trunk.

4020. By Vo Minh Thu on 2012-03-07

[FIX] signaling sequences: create them only if they do not exist.

[Vo Minh Thu (thu)]

Similar to https://code.launchpad.net/~openerp-dev/openobject-server/6.1-gunicorn-signaling-vmt/+merge/95936. Probably we can just forward port that one when it gets merged.

[Vo Minh Thu (thu)]

Was forward-ported from 6.1.

Unmerged revisions

4020. By Vo Minh Thu on 2012-03-07

[FIX] signaling sequences: create them only if they do not exist.

4019. By Vo Minh Thu on 2012-03-07

[MERGE] merged trunk.

4018. By Vo Minh Thu on 2012-03-07

[IMP] multi-process: moved signaling sequences to registry creation instead of base.sql.

=== modified file 'gunicorn.conf.py'
--- gunicorn.conf.py	2012-02-10 15:25:21 +0000
+++ gunicorn.conf.py	2012-03-07 12:45:23 +0000
@@ -25,7 +25,6 @@
 
 # Some application-wide initialization is needed.
 on_starting = openerp.wsgi.core.on_starting
-when_ready = openerp.wsgi.core.when_ready
 pre_request = openerp.wsgi.core.pre_request
 post_request = openerp.wsgi.core.post_request
 

=== modified file 'openerp/__init__.py'
--- openerp/__init__.py	2011-09-27 16:51:33 +0000
+++ openerp/__init__.py	2012-03-07 12:45:23 +0000
@@ -45,5 +45,12 @@
 import workflow
 import wsgi
 
+# Is the server running in multi-process mode (e.g. behind Gunicorn).
+# If this is True, the processes have to communicate some events,
+# e.g. database update or cache invalidation. Each process has also
+# its own copy of the data structure and we don't need to care about
+# locks between threads.
+multi_process = False
+
 # vim:expandtab:smartindent:tabstop=4:softtabstop=4:shiftwidth=4:
 

=== modified file 'openerp/addons/base/ir/ir_ui_menu.py'
--- openerp/addons/base/ir/ir_ui_menu.py	2012-02-10 08:26:37 +0000
+++ openerp/addons/base/ir/ir_ui_menu.py	2012-03-07 12:45:23 +0000
@@ -42,7 +42,7 @@
 
     def __init__(self, *args, **kwargs):
         self.cache_lock = threading.RLock()
-        self.clear_cache()
+        self._cache = {}
         r = super(ir_ui_menu, self).__init__(*args, **kwargs)
         self.pool.get('ir.model.access').register_cache_clearing_method(self._name, 'clear_cache')
         return r
@@ -50,6 +50,10 @@
     def clear_cache(self):
         with self.cache_lock:
             # radical but this doesn't frequently happen
+            if self._cache:
+                # Normally this is done by openerp.tools.ormcache
+                # but since we do not use it, set it by ourself.
+                self.pool._any_cache_cleared = True
             self._cache = {}
 
     def _filter_visible_menus(self, cr, uid, ids, context=None):

=== modified file 'openerp/addons/base/module/module.py'
--- openerp/addons/base/module/module.py	2012-02-14 15:18:46 +0000
+++ openerp/addons/base/module/module.py	2012-03-07 12:45:23 +0000
@@ -30,6 +30,7 @@
 import zipfile
 import zipimport
 
+import openerp
 import openerp.modules as addons
 import pooler
 import release
@@ -344,6 +345,7 @@
         if to_install_ids:
             self.button_install(cr, uid, to_install_ids, context=context)
 
+        openerp.modules.registry.RegistryManager.signal_registry_change(cr.dbname)
         return dict(ACTION_DICT, name=_('Install'))
 
     def button_immediate_install(self, cr, uid, ids, context=None):

=== modified file 'openerp/cron.py'
--- openerp/cron.py	2012-01-24 11:07:30 +0000
+++ openerp/cron.py	2012-03-07 12:45:23 +0000
@@ -204,9 +204,12 @@
         _logger.warning("Connection pool size (%s) is set lower than max number of cron threads (%s), "
                         "this may cause trouble if you reach that number of parallel cron tasks.",
                         db_maxconn, _thread_slots)
-    t = threading.Thread(target=runner, name="openerp.cron.master_thread")
-    t.setDaemon(True)
-    t.start()
-    _logger.debug("Master cron daemon started!")
+    if _thread_slots:
+        t = threading.Thread(target=runner, name="openerp.cron.master_thread")
+        t.setDaemon(True)
+        t.start()
+        _logger.debug("Master cron daemon started!")
+    else:
+        _logger.info("No master cron daemon (0 workers needed).")
 
 # vim:expandtab:smartindent:tabstop=4:softtabstop=4:shiftwidth=4:

=== modified file 'openerp/modules/registry.py'
--- openerp/modules/registry.py	2012-01-24 12:42:52 +0000
+++ openerp/modules/registry.py	2012-03-07 12:45:23 +0000
@@ -51,6 +51,18 @@
         self.db_name = db_name
         self.db = openerp.sql_db.db_connect(db_name)
 
+        # Inter-process signaling (used only when openerp.multi_process is True):
+        # The `base_registry_signaling` sequence indicates the whole registry
+        # must be reloaded.
+        # The `base_cache_signaling sequence` indicates all caches must be
+        # invalidated (i.e. cleared).
+        self.base_registry_signaling_sequence = 1
+        self.base_cache_signaling_sequence = 1
+
+        # Flag indicating if at least one model cache has been cleared.
+        # Useful only in a multi-process context.
+        self._any_cache_cleared = False
+
         cr = self.db.cursor()
         has_unaccent = openerp.modules.db.has_unaccent(cr)
         if openerp.tools.config['unaccent'] and not has_unaccent:
@@ -114,6 +126,36 @@
         """
         for model in self.models.itervalues():
             model.clear_caches()
+        # Special case for ir_ui_menu which does not use openerp.tools.ormcache.
+        ir_ui_menu = self.models.get('ir.ui.menu')
+        if ir_ui_menu:
+            ir_ui_menu.clear_cache()
+
+
+    # Useful only in a multi-process context.
+    def reset_any_cache_cleared(self):
+        self._any_cache_cleared = False
+
+    # Useful only in a multi-process context.
+    def any_cache_cleared(self):
+        return self._any_cache_cleared
+
+    @classmethod
+    def setup_multi_process_signaling(cls, cr):
+        if not openerp.multi_process:
+            return
+
+        # Inter-process signaling:
+        # The `base_registry_signaling` sequence indicates the whole registry
+        # must be reloaded.
+        # The `base_cache_signaling sequence` indicates all caches must be
+        # invalidated (i.e. cleared).
+        cr.execute("""SELECT sequence_name FROM information_schema.sequences WHERE sequence_name='base_registry_signaling'""")
+        if not cr.fetchall():
+            cr.execute("""CREATE SEQUENCE base_registry_signaling INCREMENT BY 1 START WITH 1""")
+            cr.execute("""SELECT nextval('base_registry_signaling')""")
+            cr.execute("""CREATE SEQUENCE base_cache_signaling INCREMENT BY 1 START WITH 1""")
+            cr.execute("""SELECT nextval('base_cache_signaling')""")
 
 class RegistryManager(object):
     """ Model registries manager.
@@ -164,6 +206,7 @@
 
             cr = registry.db.cursor()
             try:
+                Registry.setup_multi_process_signaling(cr)
                 registry.do_parent_store(cr)
                 registry.get('ir.actions.report.xml').register_all(cr)
                 cr.commit()
@@ -215,5 +258,70 @@
             if db_name in cls.registries:
                 cls.registries[db_name].clear_caches()
 
+    @classmethod
+    def check_registry_signaling(cls, db_name):
+        if openerp.multi_process and db_name in cls.registries:
+            # Check if the model registry must be reloaded (e.g. after the
+            # database has been updated by another process).
+            registry = cls.get(db_name, pooljobs=False)
+            cr = registry.db.cursor()
+            registry_reloaded = False
+            try:
+                cr.execute('SELECT last_value FROM base_registry_signaling')
+                r = cr.fetchone()[0]
+                if registry.base_registry_signaling_sequence != r:
+                    _logger.info("Reloading the model registry after database signaling.")
+                    # Don't run the cron in the Gunicorn worker.
+                    registry = cls.new(db_name, pooljobs=False)
+                    registry.base_registry_signaling_sequence = r
+                    registry_reloaded = True
+            finally:
+                cr.close()
+
+            # Check if the model caches must be invalidated (e.g. after a write
+            # occured on another process). Don't clear right after a registry
+            # has been reload.
+            cr = openerp.sql_db.db_connect(db_name).cursor()
+            try:
+                cr.execute('SELECT last_value FROM base_cache_signaling')
+                r = cr.fetchone()[0]
+                if registry.base_cache_signaling_sequence != r and not registry_reloaded:
+                    _logger.info("Invalidating all model caches after database signaling.")
+                    registry.base_cache_signaling_sequence = r
+                    registry.clear_caches()
+                    registry.reset_any_cache_cleared()
+            finally:
+                cr.close()
+
+    @classmethod
+    def signal_caches_change(cls, db_name):
+        if openerp.multi_process and db_name in cls.registries:
+            # Check the registries if any cache has been cleared and signal it
+            # through the database to other processes.
+            registry = cls.get(db_name, pooljobs=False)
+            if registry.any_cache_cleared():
+                _logger.info("At least one model cache has been cleare, signaling through the database.")
+                cr = registry.db.cursor()
+                r = 1
+                try:
+                    cr.execute("select nextval('base_cache_signaling')")
+                    r = cr.fetchone()[0]
+                finally:
+                    cr.close()
+                registry.base_cache_signaling_sequence = r
+                registry.reset_any_cache_cleared()
+
+    @classmethod
+    def signal_registry_change(cls, db_name):
+        if openerp.multi_process and db_name in cls.registries:
+            registry = cls.get(db_name, pooljobs=False)
+            cr = registry.db.cursor()
+            r = 1
+            try:
+                cr.execute("select nextval('base_registry_signaling')")
+                r = cr.fetchone()[0]
+            finally:
+                cr.close()
+            registry.base_registry_signaling_sequence = r
 
 # vim:expandtab:smartindent:tabstop=4:softtabstop=4:shiftwidth=4:

=== modified file 'openerp/osv/orm.py'
--- openerp/osv/orm.py	2012-02-15 10:17:14 +0000
+++ openerp/osv/orm.py	2012-03-07 12:45:23 +0000
@@ -2389,6 +2389,7 @@
         try:
             getattr(self, '_ormcache')
             self._ormcache = {}
+            self.pool._any_cache_cleared = True
         except AttributeError:
             pass
 

=== modified file 'openerp/service/web_services.py'
--- openerp/service/web_services.py	2012-02-14 19:30:30 +0000
+++ openerp/service/web_services.py	2012-03-07 12:45:23 +0000
@@ -568,8 +568,10 @@
             raise NameError("Method not available %s" % method)
         security.check(db,uid,passwd)
         assert openerp.osv.osv.service, "The object_proxy class must be started with start_object_proxy."
+        openerp.modules.registry.RegistryManager.check_registry_signaling(db)
         fn = getattr(openerp.osv.osv.service, method)
         res = fn(db, uid, *params)
+        openerp.modules.registry.RegistryManager.signal_caches_change(db)
         return res
 
 
@@ -649,8 +651,10 @@
         if method not in ['report', 'report_get', 'render_report']:
             raise KeyError("Method not supported %s" % method)
         security.check(db,uid,passwd)
+        openerp.modules.registry.RegistryManager.check_registry_signaling(db)
         fn = getattr(self, 'exp_' + method)
         res = fn(db, uid, *params)
+        openerp.modules.registry.RegistryManager.signal_caches_change(db)
         return res
 
     def exp_render_report(self, db, uid, object, ids, datas=None, context=None):

=== modified file 'openerp/tools/cache.py'
--- openerp/tools/cache.py	2011-11-22 08:58:48 +0000
+++ openerp/tools/cache.py	2012-03-07 12:45:23 +0000
@@ -57,10 +57,12 @@
             try:
                 key = args[self.skiparg-2:]
                 del d[key]
+                self2.pool._any_cache_cleared = True
             except KeyError:
                 pass
         else:
             d.clear()
+            self2.pool._any_cache_cleared = True
 
 class ormcache_multi(ormcache):
     def __init__(self, skiparg=2, size=8192, multi=3):

=== modified file 'openerp/wsgi/core.py'
--- openerp/wsgi/core.py	2012-02-21 18:54:41 +0000
+++ openerp/wsgi/core.py	2012-03-07 12:45:23 +0000
@@ -447,7 +447,7 @@
 
     The WSGI server can be shutdown with stop_server() below.
     """
-    threading.Thread(target=serve).start()
+    threading.Thread(name='WSGI server', target=serve).start()
 
 def stop_server():
     """ Initiate the shutdown of the WSGI server.
@@ -465,7 +465,7 @@
 def on_starting(server):
     global arbiter_pid
     arbiter_pid = os.getpid() # TODO check if this is true even after replacing the executable
-    #openerp.tools.cache = kill_workers_cache
+    openerp.multi_process = True # Yay!
     openerp.netsvc.init_logger()
     openerp.osv.osv.start_object_proxy()
     openerp.service.web_services.start_web_services()
@@ -482,11 +482,6 @@
 Maybe you forgot to add those addons in your addons_path configuration."""
             _logger.exception('Failed to load server-wide module `%s`.%s', m, msg)
 
-# Install our own signal handler on the master process.
-def when_ready(server):
-    # Hijack gunicorn's SIGWINCH handling; we can choose another one.
-    signal.signal(signal.SIGWINCH, make_winch_handler(server))
-
 # Install limits on virtual memory and CPU time consumption.
 def pre_request(worker, req):
     import os
@@ -514,30 +509,9 @@
             'too high, rebooting the worker.')
         worker.alive = False # Commit suicide after the request.
 
-# Our signal handler will signal a SGIQUIT to all workers.
-def make_winch_handler(server):
-    def handle_winch(sig, fram):
-        server.kill_workers(signal.SIGQUIT) # This is gunicorn specific.
-    return handle_winch
-
 # SIGXCPU (exceeded CPU time) signal handler will raise an exception.
 def time_expired(n, stack):
     _logger.info('CPU time limit exceeded.')
     raise Exception('CPU time limit exceeded.') # TODO one of openerp.exception
 
-# Kill gracefuly the workers (e.g. because we want to clear their cache).
-# This is done by signaling a SIGWINCH to the master process, so it can be
-# called by the workers themselves.
-def kill_workers():
-    try:
-        os.kill(arbiter_pid, signal.SIGWINCH)
-    except OSError, e:
-        if e.errno == errno.ESRCH: # no such pid
-            return
-        raise
-
-class kill_workers_cache(openerp.tools.ormcache):
-    def clear(self, dbname, *args, **kwargs):
-        kill_workers()
-
 # vim:expandtab:smartindent:tabstop=4:softtabstop=4:shiftwidth=4: