Merge pull request #1930 from minrk/limitdictdb

add size-limiting to the DictDB backend

Adds configurables for culling old records when limits are exceeded.

Default limits: 1GB of buffers, 1024 records, 10% cull.

IPython/parallel/controller/dictdb.py

@@ -1,220 +1,306 b''
 """A Task logger that presents our DB interface,
 but exists entirely in memory and implemented with dicts.
 
 Authors:
 
 * Min RK
 
 
 TaskRecords are dicts of the form:
 {
     'msg_id' : str(uuid),
     'client_uuid' : str(uuid),
     'engine_uuid' : str(uuid) or None,
     'header' : dict(header),
     'content': dict(content),
     'buffers': list(buffers),
     'submitted': datetime,
     'started': datetime or None,
     'completed': datetime or None,
     'resubmitted': datetime or None,
     'result_header' : dict(header) or None,
     'result_content' : dict(content) or None,
     'result_buffers' : list(buffers) or None,
 }
 With this info, many of the special categories of tasks can be defined by query:
 
 pending:  completed is None
 client's outstanding: client_uuid = uuid && completed is None
 MIA: arrived is None (and completed is None)
 etc.
 
 EngineRecords are dicts of the form:
 {
     'eid' : int(id),
     'uuid': str(uuid)
 }
 This may be extended, but is currently.
 
 We support a subset of mongodb operators:
     $lt,$gt,$lte,$gte,$ne,$in,$nin,$all,$mod,$exists
 """
 #-----------------------------------------------------------------------------
 #  Copyright (C) 2010-2011  The IPython Development Team
 #
 #  Distributed under the terms of the BSD License.  The full license is in
 #  the file COPYING, distributed as part of this software.
 #-----------------------------------------------------------------------------
 
 from copy import deepcopy as copy
 from datetime import datetime
 
 from IPython.config.configurable import LoggingConfigurable
 
-from IPython.utils.traitlets import Dict, Unicode, Instance
+from IPython.utils.traitlets import Dict, Unicode, Integer, Float
 
 filters = {
  '$lt' : lambda a,b: a < b,
  '$gt' : lambda a,b: b > a,
  '$eq' : lambda a,b: a == b,
  '$ne' : lambda a,b: a != b,
  '$lte': lambda a,b: a <= b,
  '$gte': lambda a,b: a >= b,
  '$in' : lambda a,b: a in b,
  '$nin': lambda a,b: a not in b,
  '$all': lambda a,b: all([ a in bb for bb in b ]),
  '$mod': lambda a,b: a%b[0] == b[1],
  '$exists' : lambda a,b: (b and a is not None) or (a is None and not b)
 }
 
 
 class CompositeFilter(object):
     """Composite filter for matching multiple properties."""
 
     def __init__(self, dikt):
         self.tests = []
         self.values = []
         for key, value in dikt.iteritems():
             self.tests.append(filters[key])
             self.values.append(value)
 
     def __call__(self, value):
         for test,check in zip(self.tests, self.values):
             if not test(value, check):
                 return False
         return True
 
 class BaseDB(LoggingConfigurable):
     """Empty Parent class so traitlets work on DB."""
     # base configurable traits:
     session = Unicode("")
 
 class DictDB(BaseDB):
     """Basic in-memory dict-based object for saving Task Records.
 
     This is the first object to present the DB interface
     for logging tasks out of memory.
 
     The interface is based on MongoDB, so adding a MongoDB
     backend should be straightforward.
     """
 
     _records = Dict()
+    _culled_ids = set() # set of ids which have been culled
+    _buffer_bytes = Integer(0) # running total of the bytes in the DB
+    
+    size_limit = Integer(1024*1024, config=True,
+        help="""The maximum total size (in bytes) of the buffers stored in the db
+        
+        When the db exceeds this size, the oldest records will be culled until
+        the total size is under size_limit * (1-cull_fraction).
+        """
+    )
+    record_limit = Integer(1024, config=True,
+        help="""The maximum number of records in the db
+        
+        When the history exceeds this size, the first record_limit * cull_fraction
+        records will be culled.
+        """
+    )
+    cull_fraction = Float(0.1, config=True,
+        help="""The fraction by which the db should culled when one of the limits is exceeded
+        
+        In general, the db size will spend most of its time with a size in the range:
+        
+        [limit * (1-cull_fraction), limit]
+        
+        for each of size_limit and record_limit.
+        """
+    )
 
     def _match_one(self, rec, tests):
         """Check if a specific record matches tests."""
         for key,test in tests.iteritems():
             if not test(rec.get(key, None)):
                 return False
         return True
 
     def _match(self, check):
         """Find all the matches for a check dict."""
         matches = []
         tests = {}
         for k,v in check.iteritems():
             if isinstance(v, dict):
                 tests[k] = CompositeFilter(v)
             else:
                 tests[k] = lambda o: o==v
 
         for rec in self._records.itervalues():
             if self._match_one(rec, tests):
                 matches.append(copy(rec))
         return matches
 
     def _extract_subdict(self, rec, keys):
         """extract subdict of keys"""
         d = {}
         d['msg_id'] = rec['msg_id']
         for key in keys:
             d[key] = rec[key]
         return copy(d)
+    
+    # methods for monitoring size / culling history
+    
+    def _add_bytes(self, rec):
+        for key in ('buffers', 'result_buffers'):
+            for buf in rec.get(key) or []:
+                self._buffer_bytes += len(buf)
+        
+        self._maybe_cull()
+    
+    def _drop_bytes(self, rec):
+        for key in ('buffers', 'result_buffers'):
+            for buf in rec.get(key) or []:
+                self._buffer_bytes -= len(buf)
+    
+    def _cull_oldest(self, n=1):
+        """cull the oldest N records"""
+        for msg_id in self.get_history()[:n]:
+            self.log.debug("Culling record: %r", msg_id)
+            self._culled_ids.add(msg_id)
+            self.drop_record(msg_id)
+    
+    def _maybe_cull(self):
+        # cull by count:
+        if len(self._records) > self.record_limit:
+            to_cull = int(self.cull_fraction * self.record_limit)
+            self.log.info("%i records exceeds limit of %i, culling oldest %i",
+                len(self._records), self.record_limit, to_cull
+            )
+            self._cull_oldest(to_cull)
+        
+        # cull by size:
+        if self._buffer_bytes > self.size_limit:
+            limit = self.size_limit * (1 - self.cull_fraction)
+            
+            before = self._buffer_bytes
+            before_count = len(self._records)
+            culled = 0
+            while self._buffer_bytes > limit:
+                self._cull_oldest(1)
+                culled += 1
+        
+            self.log.info("%i records with total buffer size %i exceeds limit: %i. Culled oldest %i records.",
+                before_count, before, self.size_limit, culled
+            )
+    
+    # public API methods:
 
     def add_record(self, msg_id, rec):
         """Add a new Task Record, by msg_id."""
         if msg_id in self._records:
             raise KeyError("Already have msg_id %r"%(msg_id))
         self._records[msg_id] = rec
+        self._add_bytes(rec)
+        self._maybe_cull()
 
     def get_record(self, msg_id):
         """Get a specific Task Record, by msg_id."""
+        if msg_id in self._culled_ids:
+            raise KeyError("Record %r has been culled for size" % msg_id)
         if not msg_id in self._records:
             raise KeyError("No such msg_id %r"%(msg_id))
         return copy(self._records[msg_id])
 
     def update_record(self, msg_id, rec):
         """Update the data in an existing record."""
-        self._records[msg_id].update(rec)
+        if msg_id in self._culled_ids:
+            raise KeyError("Record %r has been culled for size" % msg_id)
+        _rec = self._records[msg_id]
+        self._drop_bytes(_rec)
+        _rec.update(rec)
+        self._add_bytes(_rec)
 
     def drop_matching_records(self, check):
         """Remove a record from the DB."""
         matches = self._match(check)
-        for m in matches:
-            del self._records[m['msg_id']]
+        for rec in matches:
+            self._drop_bytes(rec)
+            del self._records[rec['msg_id']]
 
     def drop_record(self, msg_id):
         """Remove a record from the DB."""
+        rec = self._records[msg_id]
+        self._drop_bytes(rec)
         del self._records[msg_id]
 
-
     def find_records(self, check, keys=None):
         """Find records matching a query dict, optionally extracting subset of keys.
 
         Returns dict keyed by msg_id of matching records.
 
         Parameters
         ----------
 
         check: dict
             mongodb-style query argument
         keys: list of strs [optional]
             if specified, the subset of keys to extract.  msg_id will *always* be
             included.
         """
         matches = self._match(check)
         if keys:
             return [ self._extract_subdict(rec, keys) for rec in matches ]
         else:
             return matches
 
-
     def get_history(self):
         """get all msg_ids, ordered by time submitted."""
         msg_ids = self._records.keys()
         return sorted(msg_ids, key=lambda m: self._records[m]['submitted'])
 
+
 NODATA = KeyError("NoDB backend doesn't store any data. "
 "Start the Controller with a DB backend to enable resubmission / result persistence."
 )
 
-class NoDB(DictDB):
+
+class NoDB(BaseDB):
     """A blackhole db backend that actually stores no information.
     
     Provides the full DB interface, but raises KeyErrors on any
     method that tries to access the records.  This can be used to
     minimize the memory footprint of the Hub when its record-keeping
     functionality is not required.
     """
     
     def add_record(self, msg_id, record):
         pass
     
     def get_record(self, msg_id):
         raise NODATA
     
     def update_record(self, msg_id, record):
         pass
     
     def drop_matching_records(self, check):
         pass
     
     def drop_record(self, msg_id):
         pass
     
     def find_records(self, check, keys=None):
         raise NODATA
     
     def get_history(self):
         raise NODATA
 

IPython/parallel/tests/test_db.py

@@ -1,249 +1,314 b''
 """Tests for db backends
 
 Authors:
 
 * Min RK
 """
 
 #-------------------------------------------------------------------------------
 #  Copyright (C) 2011  The IPython Development Team
 #
 #  Distributed under the terms of the BSD License.  The full license is in
 #  the file COPYING, distributed as part of this software.
 #-------------------------------------------------------------------------------
 
 #-------------------------------------------------------------------------------
 # Imports
 #-------------------------------------------------------------------------------
 
 from __future__ import division
 
 import logging
 import os
 import tempfile
 import time
 
 from datetime import datetime, timedelta
 from unittest import TestCase
 
 from IPython.parallel import error
 from IPython.parallel.controller.dictdb import DictDB
 from IPython.parallel.controller.sqlitedb import SQLiteDB
 from IPython.parallel.controller.hub import init_record, empty_record
 
 from IPython.testing import decorators as dec
 from IPython.zmq.session import Session
 
 
 #-------------------------------------------------------------------------------
 # TestCases
 #-------------------------------------------------------------------------------
 
 
 def setup():
     global temp_db
     temp_db = tempfile.NamedTemporaryFile(suffix='.db').name
 
 
-class TestDictBackend(TestCase):
+class TaskDBTest:
     def setUp(self):
         self.session = Session()
         self.db = self.create_db()
         self.load_records(16)
     
     def create_db(self):
-        return DictDB()
+        raise NotImplementedError
     
-    def load_records(self, n=1):
+    def load_records(self, n=1, buffer_size=100):
         """load n records for testing"""
         #sleep 1/10 s, to ensure timestamp is different to previous calls
         time.sleep(0.1)
         msg_ids = []
         for i in range(n):
             msg = self.session.msg('apply_request', content=dict(a=5))
-            msg['buffers'] = []
+            msg['buffers'] = [os.urandom(buffer_size)]
             rec = init_record(msg)
             msg_id = msg['header']['msg_id']
             msg_ids.append(msg_id)
             self.db.add_record(msg_id, rec)
         return msg_ids
     
     def test_add_record(self):
         before = self.db.get_history()
         self.load_records(5)
         after = self.db.get_history()
         self.assertEqual(len(after), len(before)+5)
         self.assertEqual(after[:-5],before)
         
     def test_drop_record(self):
         msg_id = self.load_records()[-1]
         rec = self.db.get_record(msg_id)
         self.db.drop_record(msg_id)
         self.assertRaises(KeyError,self.db.get_record, msg_id)
     
     def _round_to_millisecond(self, dt):
         """necessary because mongodb rounds microseconds"""
         micro = dt.microsecond
         extra = int(str(micro)[-3:])
         return dt - timedelta(microseconds=extra)
     
     def test_update_record(self):
         now = self._round_to_millisecond(datetime.now())
         # 
         msg_id = self.db.get_history()[-1]
         rec1 = self.db.get_record(msg_id)
         data = {'stdout': 'hello there', 'completed' : now}
         self.db.update_record(msg_id, data)
         rec2 = self.db.get_record(msg_id)
         self.assertEqual(rec2['stdout'], 'hello there')
         self.assertEqual(rec2['completed'], now)
         rec1.update(data)
         self.assertEqual(rec1, rec2)
     
     # def test_update_record_bad(self):
     #     """test updating nonexistant records"""
     #     msg_id = str(uuid.uuid4())
     #     data = {'stdout': 'hello there'}
     #     self.assertRaises(KeyError, self.db.update_record, msg_id, data)
 
     def test_find_records_dt(self):
         """test finding records by date"""
         hist = self.db.get_history()
         middle = self.db.get_record(hist[len(hist)//2])
         tic = middle['submitted']
         before = self.db.find_records({'submitted' : {'$lt' : tic}})
         after = self.db.find_records({'submitted' : {'$gte' : tic}})
         self.assertEqual(len(before)+len(after),len(hist))
         for b in before:
             self.assertTrue(b['submitted'] < tic)
         for a in after:
             self.assertTrue(a['submitted'] >= tic)
         same = self.db.find_records({'submitted' : tic})
         for s in same:
             self.assertTrue(s['submitted'] == tic)
     
     def test_find_records_keys(self):
         """test extracting subset of record keys"""
         found = self.db.find_records({'msg_id': {'$ne' : ''}},keys=['submitted', 'completed'])
         for rec in found:
             self.assertEqual(set(rec.keys()), set(['msg_id', 'submitted', 'completed']))
     
     def test_find_records_msg_id(self):
         """ensure msg_id is always in found records"""
         found = self.db.find_records({'msg_id': {'$ne' : ''}},keys=['submitted', 'completed'])
         for rec in found:
             self.assertTrue('msg_id' in rec.keys())
         found = self.db.find_records({'msg_id': {'$ne' : ''}},keys=['submitted'])
         for rec in found:
             self.assertTrue('msg_id' in rec.keys())
         found = self.db.find_records({'msg_id': {'$ne' : ''}},keys=['msg_id'])
         for rec in found:
             self.assertTrue('msg_id' in rec.keys())
     
     def test_find_records_in(self):
         """test finding records with '$in','$nin' operators"""
         hist = self.db.get_history()
         even = hist[::2]
         odd = hist[1::2]
         recs = self.db.find_records({ 'msg_id' : {'$in' : even}})
         found = [ r['msg_id'] for r in recs ]
         self.assertEqual(set(even), set(found))
         recs = self.db.find_records({ 'msg_id' : {'$nin' : even}})
         found = [ r['msg_id'] for r in recs ]
         self.assertEqual(set(odd), set(found))
     
     def test_get_history(self):
         msg_ids = self.db.get_history()
         latest = datetime(1984,1,1)
         for msg_id in msg_ids:
             rec = self.db.get_record(msg_id)
             newt = rec['submitted']
             self.assertTrue(newt >= latest)
             latest = newt
         msg_id = self.load_records(1)[-1]
         self.assertEqual(self.db.get_history()[-1],msg_id)
     
     def test_datetime(self):
         """get/set timestamps with datetime objects"""
         msg_id = self.db.get_history()[-1]
         rec = self.db.get_record(msg_id)
         self.assertTrue(isinstance(rec['submitted'], datetime))
         self.db.update_record(msg_id, dict(completed=datetime.now()))
         rec = self.db.get_record(msg_id)
         self.assertTrue(isinstance(rec['completed'], datetime))
 
     def test_drop_matching(self):
         msg_ids = self.load_records(10)
         query = {'msg_id' : {'$in':msg_ids}}
         self.db.drop_matching_records(query)
         recs = self.db.find_records(query)
         self.assertEqual(len(recs), 0)
     
     def test_null(self):
         """test None comparison queries"""
         msg_ids = self.load_records(10)
 
         query = {'msg_id' : None}
         recs = self.db.find_records(query)
         self.assertEqual(len(recs), 0)
 
         query = {'msg_id' : {'$ne' : None}}
         recs = self.db.find_records(query)
         self.assertTrue(len(recs) >= 10)
     
     def test_pop_safe_get(self):
         """editing query results shouldn't affect record [get]"""
         msg_id = self.db.get_history()[-1]
         rec = self.db.get_record(msg_id)
         rec.pop('buffers')
         rec['garbage'] = 'hello'
         rec['header']['msg_id'] = 'fubar'
         rec2 = self.db.get_record(msg_id)
         self.assertTrue('buffers' in rec2)
         self.assertFalse('garbage' in rec2)
         self.assertEqual(rec2['header']['msg_id'], msg_id)
     
     def test_pop_safe_find(self):
         """editing query results shouldn't affect record [find]"""
         msg_id = self.db.get_history()[-1]
         rec = self.db.find_records({'msg_id' : msg_id})[0]
         rec.pop('buffers')
         rec['garbage'] = 'hello'
         rec['header']['msg_id'] = 'fubar'
         rec2 = self.db.find_records({'msg_id' : msg_id})[0]
         self.assertTrue('buffers' in rec2)
         self.assertFalse('garbage' in rec2)
         self.assertEqual(rec2['header']['msg_id'], msg_id)
 
     def test_pop_safe_find_keys(self):
         """editing query results shouldn't affect record [find+keys]"""
         msg_id = self.db.get_history()[-1]
         rec = self.db.find_records({'msg_id' : msg_id}, keys=['buffers', 'header'])[0]
         rec.pop('buffers')
         rec['garbage'] = 'hello'
         rec['header']['msg_id'] = 'fubar'
         rec2 = self.db.find_records({'msg_id' : msg_id})[0]
         self.assertTrue('buffers' in rec2)
         self.assertFalse('garbage' in rec2)
         self.assertEqual(rec2['header']['msg_id'], msg_id)
 
 
-class TestSQLiteBackend(TestDictBackend):
+class TestDictBackend(TaskDBTest, TestCase):
+    
+    def create_db(self):
+        return DictDB()
+    
+    def test_cull_count(self):
+        self.db = self.create_db() # skip the load-records init from setUp
+        self.db.record_limit = 20
+        self.db.cull_fraction = 0.2
+        self.load_records(20)
+        self.assertEquals(len(self.db.get_history()), 20)
+        self.load_records(1)
+        # 0.2 * 20 = 4, 21 - 4 = 17
+        self.assertEquals(len(self.db.get_history()), 17)
+        self.load_records(3)
+        self.assertEquals(len(self.db.get_history()), 20)
+        self.load_records(1)
+        self.assertEquals(len(self.db.get_history()), 17)
+        
+        for i in range(100):
+            self.load_records(1)
+            self.assertTrue(len(self.db.get_history()) >= 17)
+            self.assertTrue(len(self.db.get_history()) <= 20)
+
+    def test_cull_size(self):
+        self.db = self.create_db() # skip the load-records init from setUp
+        self.db.size_limit = 1000
+        self.db.cull_fraction = 0.2
+        self.load_records(100, buffer_size=10)
+        self.assertEquals(len(self.db.get_history()), 100)
+        self.load_records(1, buffer_size=0)
+        self.assertEquals(len(self.db.get_history()), 101)
+        self.load_records(1, buffer_size=1)
+        # 0.2 * 100 = 20, 101 - 20 = 81
+        self.assertEquals(len(self.db.get_history()), 81)
+    
+    def test_cull_size_drop(self):
+        """dropping records updates tracked buffer size"""
+        self.db = self.create_db() # skip the load-records init from setUp
+        self.db.size_limit = 1000
+        self.db.cull_fraction = 0.2
+        self.load_records(100, buffer_size=10)
+        self.assertEquals(len(self.db.get_history()), 100)
+        self.db.drop_record(self.db.get_history()[-1])
+        self.assertEquals(len(self.db.get_history()), 99)
+        self.load_records(1, buffer_size=5)
+        self.assertEquals(len(self.db.get_history()), 100)
+        self.load_records(1, buffer_size=5)
+        self.assertEquals(len(self.db.get_history()), 101)
+        self.load_records(1, buffer_size=1)
+        self.assertEquals(len(self.db.get_history()), 81)
+
+    def test_cull_size_update(self):
+        """updating records updates tracked buffer size"""
+        self.db = self.create_db() # skip the load-records init from setUp
+        self.db.size_limit = 1000
+        self.db.cull_fraction = 0.2
+        self.load_records(100, buffer_size=10)
+        self.assertEquals(len(self.db.get_history()), 100)
+        msg_id = self.db.get_history()[-1]
+        self.db.update_record(msg_id, dict(result_buffers = [os.urandom(10)], buffers=[]))
+        self.assertEquals(len(self.db.get_history()), 100)
+        self.db.update_record(msg_id, dict(result_buffers = [os.urandom(11)], buffers=[]))
+        self.assertEquals(len(self.db.get_history()), 79)
+
+class TestSQLiteBackend(TaskDBTest, TestCase):
 
     @dec.skip_without('sqlite3')
     def create_db(self):
         location, fname = os.path.split(temp_db)
         log = logging.getLogger('test')
         log.setLevel(logging.CRITICAL)
         return SQLiteDB(location=location, fname=fname, log=log)
     
     def tearDown(self):
         self.db._db.close()
 
 
 def teardown():
     """cleanup task db file after all tests have run"""
     try:
         os.remove(temp_db)
     except:
         pass

IPython/parallel/tests/test_mongodb.py

@@ -1,42 +1,44 b''
 """Tests for mongodb backend
 
 Authors:
 
 * Min RK
 """
 
 #-------------------------------------------------------------------------------
 #  Copyright (C) 2011  The IPython Development Team
 #
 #  Distributed under the terms of the BSD License.  The full license is in
 #  the file COPYING, distributed as part of this software.
 #-------------------------------------------------------------------------------
 
 #-------------------------------------------------------------------------------
 # Imports
 #-------------------------------------------------------------------------------
 
+from unittest import TestCase
+
 from nose import SkipTest
 
 from pymongo import Connection
 from IPython.parallel.controller.mongodb import MongoDB
 
 from . import test_db
 
 try:
     c = Connection()
 except Exception:
     c=None
 
-class TestMongoBackend(test_db.TestDictBackend):
+class TestMongoBackend(test_db.TaskDBTest, TestCase):
     """MongoDB backend tests"""
 
     def create_db(self):
         try:
             return MongoDB(database='iptestdb', _connection=c)
         except Exception:
             raise SkipTest("Couldn't connect to mongodb")
 
 def teardown(self):
     if c is not None:
         c.drop_database('iptestdb')