]> git.openstreetmap.org Git - nominatim.git/blobdiff - nominatim/nominatim.py
indexer: allow batch processing of places
[nominatim.git] / nominatim / nominatim.py
old mode 100644 (file)
new mode 100755 (executable)
index 610e3de..e8600ca
@@ -1,4 +1,4 @@
-#! /usr/bin/env python
+#! /usr/bin/env python3
 #-----------------------------------------------------------------------------
 # nominatim - [description]
 #-----------------------------------------------------------------------------
@@ -32,36 +32,136 @@ import psycopg2
 from psycopg2.extras import wait_select
 import select
 
+from indexer.progress import ProgressLogger
+
 log = logging.getLogger()
 
 def make_connection(options, asynchronous=False):
-    return psycopg2.connect(dbname=options.dbname, user=options.user,
-                            password=options.password, host=options.host,
-                            port=options.port, async_=asynchronous)
+    params = {'dbname' : options.dbname,
+              'user' : options.user,
+              'password' : options.password,
+              'host' : options.host,
+              'port' : options.port,
+              'async' : asynchronous}
+
+    return psycopg2.connect(**params)
 
-class IndexingThread(object):
 
-    def __init__(self, thread_num, options):
-        log.debug("Creating thread {}".format(thread_num))
-        self.thread_num = thread_num
+class RankRunner(object):
+    """ Returns SQL commands for indexing one rank within the placex table.
+    """
+
+    def __init__(self, rank):
+        self.rank = rank
+
+    def name(self):
+        return "rank {}".format(self.rank)
+
+    def sql_count_objects(self):
+        return """SELECT count(*) FROM placex
+                  WHERE rank_search = {} and indexed_status > 0
+               """.format(self.rank)
+
+    def sql_get_objects(self):
+        return """SELECT place_id FROM placex
+                  WHERE indexed_status > 0 and rank_search = {}
+                  ORDER BY geometry_sector""".format(self.rank)
+
+    def sql_index_place(self, ids):
+        return "UPDATE placex SET indexed_status = 0 WHERE place_id IN ({})"\
+               .format(','.join((str(i) for i in ids)))
+
+
+class InterpolationRunner(object):
+    """ Returns SQL commands for indexing the address interpolation table
+        location_property_osmline.
+    """
+
+    def name(self):
+        return "interpolation lines (location_property_osmline)"
+
+    def sql_count_objects(self):
+        return """SELECT count(*) FROM location_property_osmline
+                  WHERE indexed_status > 0"""
+
+    def sql_get_objects(self):
+        return """SELECT place_id FROM location_property_osmline
+                  WHERE indexed_status > 0
+                  ORDER BY geometry_sector"""
+
+    def sql_index_place(self, ids):
+        return """UPDATE location_property_osmline
+                  SET indexed_status = 0 WHERE place_id IN ({})"""\
+               .format(','.join((str(i) for i in ids)))
+
+
+class DBConnection(object):
+    """ A single non-blocking database connection.
+    """
+
+    def __init__(self, options):
+        self.current_query = None
+        self.current_params = None
+
+        self.conn = None
+        self.connect()
+
+    def connect(self):
+        if self.conn is not None:
+            self.cursor.close()
+            self.conn.close()
+
         self.conn = make_connection(options, asynchronous=True)
         self.wait()
 
         self.cursor = self.conn.cursor()
-        self.perform("SET lc_messages TO 'C'")
+        # Disable JIT and parallel workers as they are known to cause problems.
+        # Update pg_settings instead of using SET because it does not yield
+        # errors on older versions of Postgres where the settings are not
+        # implemented.
+        self.perform(
+            """ UPDATE pg_settings SET setting = -1 WHERE name = 'jit_above_cost';
+                UPDATE pg_settings SET setting = 0 
+                   WHERE name = 'max_parallel_workers_per_gather';""")
         self.wait()
 
-        self.current_query = None
-
     def wait(self):
-        wait_select(self.conn)
-        self.current_query = None
+        """ Block until any pending operation is done.
+        """
+        while True:
+            try:
+                wait_select(self.conn)
+                self.current_query = None
+                return
+            except psycopg2.extensions.TransactionRollbackError as e:
+                if e.pgcode == '40P01':
+                    log.info("Deadlock detected (params = {}), retry."
+                              .format(self.current_params))
+                    self.cursor.execute(self.current_query, self.current_params)
+                else:
+                    raise
+            except psycopg2.errors.DeadlockDetected:
+                self.cursor.execute(self.current_query, self.current_params)
 
     def perform(self, sql, args=None):
+        """ Send SQL query to the server. Returns immediately without
+            blocking.
+        """
         self.current_query = sql
+        self.current_params = args
         self.cursor.execute(sql, args)
 
+    def fileno(self):
+        """ File descriptor to wait for. (Makes this class select()able.)
+        """
+        return self.conn.fileno()
+
     def is_done(self):
+        """ Check if the connection is available for a new query.
+
+            Also checks if the previous query has run into a deadlock.
+            If so, then the previous query is repeated.
+        """
         if self.current_query is None:
             return True
 
@@ -70,157 +170,106 @@ class IndexingThread(object):
                 self.current_query = None
                 return True
         except psycopg2.extensions.TransactionRollbackError as e:
-            if e.pgcode is None:
-                raise RuntimeError("Postgres exception has no error code")
             if e.pgcode == '40P01':
-                log.info("Deadlock detected, retry.")
-                self.cursor.execute(sql)
+                log.info("Deadlock detected (params = {}), retry.".format(self.current_params))
+                self.cursor.execute(self.current_query, self.current_params)
             else:
                 raise
+        except psycopg2.errors.DeadlockDetected:
+            self.cursor.execute(self.current_query, self.current_params)
 
+        return False
 
 
 class Indexer(object):
+    """ Main indexing routine.
+    """
 
     def __init__(self, options):
-        self.options = options
+        self.minrank = max(0, options.minrank)
+        self.maxrank = min(30, options.maxrank)
         self.conn = make_connection(options)
-
-        self.threads = []
-        self.poll = select.poll()
-        for i in range(options.threads):
-            t = IndexingThread(i, options)
-            self.threads.append(t)
-            self.poll.register(t.conn.fileno(), select.EPOLLIN)
-        self.next_thread = 0
+        self.threads = [DBConnection(options) for i in range(options.threads)]
 
     def run(self):
-        log.info("Starting indexing rank ({} to {}) using {} threads".format(
-                 self.options.minrank, self.options.maxrank,
-                 self.options.threads))
+        """ Run indexing over the entire database.
+        """
+        log.warning("Starting indexing rank ({} to {}) using {} threads".format(
+                 self.minrank, self.maxrank, len(self.threads)))
 
-        for rank in range(self.options.minrank, 30):
+        for rank in range(self.minrank, self.maxrank):
             self.index(RankRunner(rank))
 
-        if self.options.maxrank >= 30:
-            self.index(InterpolationRunner())
-            self.index(RankRunner(30))
+        if self.maxrank == 30:
+            self.index(InterpolationRunner(), 20)
 
-    def index(self, obj):
-        log.info("Starting {}".format(obj.name()))
+        self.index(RankRunner(self.maxrank), 20)
 
-        cur = self.conn.cursor(name="main")
-        cur.execute(obj.sql_index_sectors())
+    def index(self, obj, batch=1):
+        """ Index a single rank or table. `obj` describes the SQL to use
+            for indexing. `batch` describes the number of objects that
+            should be processed with a single SQL statement
+        """
+        log.warning("Starting {}".format(obj.name()))
 
-        total_tuples = 0
-        for r in cur:
-            total_tuples += r[1]
-        log.debug("Total number of rows; {}".format(total_tuples))
+        cur = self.conn.cursor()
+        cur.execute(obj.sql_count_objects())
 
-        cur.scroll(0, mode='absolute')
+        total_tuples = cur.fetchone()[0]
+        log.debug("Total number of rows: {}".format(total_tuples))
 
-        done_tuples = 0
-        rank_start_time = datetime.now()
-        for r in cur:
-            sector = r[0]
-
-            # Should we do the remaining ones together?
-            do_all = total_tuples - done_tuples < len(self.threads) * 1000
-
-            pcur = self.conn.cursor(name='places')
+        cur.close()
 
-            if do_all:
-                pcur.execute(obj.sql_nosector_places())
-            else:
-                pcur.execute(obj.sql_sector_places(), (sector, ))
+        next_thread = self.find_free_thread()
+        progress = ProgressLogger(obj.name(), total_tuples)
 
-            for place in pcur:
-                place_id = place[0]
-                log.debug("Processing place {}".format(place_id))
-                thread = self.find_free_thread()
+        cur = self.conn.cursor(name='places')
+        cur.execute(obj.sql_get_objects())
 
-                thread.perform(obj.sql_index_place(), (place_id,))
-                done_tuples += 1
+        while True:
+            places = [p[0] for p in cur.fetchmany(batch)]
+            if len(places) == 0:
+                break
 
-            pcur.close()
+            log.debug("Processing places: {}".format(places))
+            thread = next(next_thread)
 
-            if do_all:
-                break
+            thread.perform(obj.sql_index_place(places))
+            progress.add(len(places))
 
         cur.close()
 
         for t in self.threads:
             t.wait()
 
-        rank_end_time = datetime.now()
-        diff_seconds = (rank_end_time-rank_start_time).total_seconds()
-
-        log.info("Done {} in {} @ {} per second - FINISHED {}\n".format(
-                 done_tuples, int(diff_seconds),
-                 done_tuples/diff_seconds, obj.name()))
+        progress.done()
 
     def find_free_thread(self):
-        while True:
-            for t in self.threads:
-                if t.is_done():
-                    return t
-
-            self.poll.poll()
-
-        assert(False, "Unreachable code")
-
-class RankRunner(object):
-
-    def __init__(self, rank):
-        self.rank = rank
-
-    def name(self):
-        return "rank {}".format(self.rank)
-
-    def sql_index_sectors(self):
-        return """SELECT geometry_sector, count(*) FROM placex
-                  WHERE rank_search = {} and indexed_status > 0
-                  GROUP BY geometry_sector
-                  ORDER BY geometry_sector""".format(self.rank)
-
-    def sql_nosector_places(self):
-        return """SELECT place_id FROM placex
-                  WHERE indexed_status > 0 and rank_search = {}
-                  ORDER BY geometry_sector""".format(self.rank)
-
-    def sql_sector_places(self):
-        return """SELECT place_id FROM placex
-                  WHERE indexed_status > 0 and geometry_sector = %s
-                  ORDER BY geometry_sector"""
-
-    def sql_index_place(self):
-        return "UPDATE placex SET indexed_status = 0 WHERE place_id = %s"
-
-
-class InterpolationRunner(object):
-
-    def name(self):
-        return "interpolation lines (location_property_osmline)"
+        """ Generator that returns the next connection that is free for
+            sending a query.
+        """
+        ready = self.threads
+        command_stat = 0
 
-    def sql_index_sectors(self):
-        return """SELECT geometry_sector, count(*) FROM location_property_osmline
-                  WHERE indexed_status > 0
-                  GROUP BY geometry_sector
-                  ORDER BY geometry_sector"""
-
-    def sql_nosector_places(self):
-        return """SELECT place_id FROM location_property_osmline
-                  WHERE indexed_status > 0
-                  ORDER BY geometry_sector"""
-
-    def sql_sector_places(self):
-        return """SELECT place_id FROM location_property_osmline
-                  WHERE indexed_status > 0 and geometry_sector = %s
-                  ORDER BY geometry_sector"""
+        while True:
+            for thread in ready:
+                if thread.is_done():
+                    command_stat += 1
+                    yield thread
+
+            # refresh the connections occasionaly to avoid potential
+            # memory leaks in Postgresql.
+            if command_stat > 100000:
+                for t in self.threads:
+                    while not t.is_done():
+                        t.wait()
+                    t.connect()
+                command_stat = 0
+                ready = self.threads
+            else:
+                ready, _, _ = select.select(self.threads, [], [])
 
-    def sql_index_place(self):
-        return """UPDATE location_property_osmline
-                  SET indexed_status = 0 WHERE place_id = %s"""
+        assert False, "Unreachable code"
 
 
 def nominatim_arg_parser():
@@ -229,7 +278,7 @@ def nominatim_arg_parser():
     def h(s):
         return re.sub("\s\s+" , " ", s)
 
-    p = ArgumentParser(description=__doc__,
+    p = ArgumentParser(description="Indexing tool for Nominatim.",
                        formatter_class=RawDescriptionHelpFormatter)
 
     p.add_argument('-d', '--database',