]> git.openstreetmap.org Git - nominatim.git/blobdiff - nominatim/indexer/indexer.py
don't even try heavily penalized searches
[nominatim.git] / nominatim / indexer / indexer.py
index 3a39a151dd4f1f1c845ee15ae1c1591a75cd0809..233423f03c6a202ec088cfeb0fe7ac26c79db01f 100644 (file)
@@ -1,77 +1,95 @@
+# SPDX-License-Identifier: GPL-2.0-only
+#
+# This file is part of Nominatim. (https://nominatim.org)
+#
+# Copyright (C) 2022 by the Nominatim developer community.
+# For a full list of authors see the git log.
 """
 Main work horse for indexing (computing addresses) the database.
 """
+from typing import Optional, Any, cast
 import logging
-import select
+import time
 
+import psycopg2.extras
+
+from nominatim.tokenizer.base import AbstractTokenizer
 from nominatim.indexer.progress import ProgressLogger
 from nominatim.indexer import runners
-from nominatim.db.async_connection import DBConnection
-from nominatim.db.connection import connect
+from nominatim.db.async_connection import DBConnection, WorkerPool
+from nominatim.db.connection import connect, Connection, Cursor
+from nominatim.typing import DictCursorResults
 
 LOG = logging.getLogger()
 
-class WorkerPool:
-    """ A pool of asynchronous database connections.
 
-        The pool may be used as a context manager.
+class PlaceFetcher:
+    """ Asynchronous connection that fetches place details for processing.
     """
-    REOPEN_CONNECTIONS_AFTER = 100000
+    def __init__(self, dsn: str, setup_conn: Connection) -> None:
+        self.wait_time = 0.0
+        self.current_ids: Optional[DictCursorResults] = None
+        self.conn: Optional[DBConnection] = DBConnection(dsn,
+                                               cursor_factory=psycopg2.extras.DictCursor)
+
+        with setup_conn.cursor() as cur:
+            # need to fetch those manually because register_hstore cannot
+            # fetch them on an asynchronous connection below.
+            hstore_oid = cur.scalar("SELECT 'hstore'::regtype::oid")
+            hstore_array_oid = cur.scalar("SELECT 'hstore[]'::regtype::oid")
+
+        psycopg2.extras.register_hstore(self.conn.conn, oid=hstore_oid,
+                                        array_oid=hstore_array_oid)
+
+    def close(self) -> None:
+        """ Close the underlying asynchronous connection.
+        """
+        if self.conn:
+            self.conn.close()
+            self.conn = None
 
-    def __init__(self, dsn, pool_size):
-        self.threads = [DBConnection(dsn) for _ in range(pool_size)]
-        self.free_workers = self._yield_free_worker()
 
+    def fetch_next_batch(self, cur: Cursor, runner: runners.Runner) -> bool:
+        """ Send a request for the next batch of places.
+            If details for the places are required, they will be fetched
+            asynchronously.
 
-    def finish_all(self):
-        """ Wait for all connection to finish.
+            Returns true if there is still data available.
         """
-        for thread in self.threads:
-            while not thread.is_done():
-                thread.wait()
+        ids = cast(Optional[DictCursorResults], cur.fetchmany(100))
 
-        self.free_workers = self._yield_free_worker()
+        if not ids:
+            self.current_ids = None
+            return False
 
-    def close(self):
-        """ Close all connections and clear the pool.
-        """
-        for thread in self.threads:
-            thread.close()
-        self.threads = []
-        self.free_workers = None
+        assert self.conn is not None
+        self.current_ids = runner.get_place_details(self.conn, ids)
 
+        return True
 
-    def next_free_worker(self):
-        """ Get the next free connection.
+    def get_batch(self) -> DictCursorResults:
+        """ Get the next batch of data, previously requested with
+            `fetch_next_batch`.
         """
-        return next(self.free_workers)
-
-
-    def _yield_free_worker(self):
-        ready = self.threads
-        command_stat = 0
-        while True:
-            for thread in ready:
-                if thread.is_done():
-                    command_stat += 1
-                    yield thread
-
-            if command_stat > self.REOPEN_CONNECTIONS_AFTER:
-                for thread in self.threads:
-                    while not thread.is_done():
-                        thread.wait()
-                    thread.connect()
-                ready = self.threads
-                command_stat = 0
-            else:
-                _, ready, _ = select.select([], self.threads, [])
-
-
-    def __enter__(self):
+        assert self.conn is not None
+        assert self.conn.cursor is not None
+
+        if self.current_ids is not None and not self.current_ids:
+            tstart = time.time()
+            self.conn.wait()
+            self.wait_time += time.time() - tstart
+            self.current_ids = cast(Optional[DictCursorResults],
+                                    self.conn.cursor.fetchall())
+
+        return self.current_ids if self.current_ids is not None else []
+
+    def __enter__(self) -> 'PlaceFetcher':
         return self
 
 
-    def __exit__(self, exc_type, exc_value, traceback):
+    def __exit__(self, exc_type: Any, exc_value: Any, traceback: Any) -> None:
+        assert self.conn is not None
+        self.conn.wait()
         self.close()
 
 
@@ -79,14 +97,25 @@ class Indexer:
     """ Main indexing routine.
     """
 
-    def __init__(self, dsn, tokenizer, num_threads):
+    def __init__(self, dsn: str, tokenizer: AbstractTokenizer, num_threads: int):
         self.dsn = dsn
         self.tokenizer = tokenizer
         self.num_threads = num_threads
 
 
-    def index_full(self, analyse=True):
-        """ Index the complete database. This will first index boudnaries
+    def has_pending(self) -> bool:
+        """ Check if any data still needs indexing.
+            This function must only be used after the import has finished.
+            Otherwise it will be very expensive.
+        """
+        with connect(self.dsn) as conn:
+            with conn.cursor() as cur:
+                cur.execute("SELECT 'a' FROM placex WHERE indexed_status > 0 LIMIT 1")
+                return cur.rowcount > 0
+
+
+    def index_full(self, analyse: bool = True) -> None:
+        """ Index the complete database. This will first index boundaries
             followed by all other objects. When `analyse` is True, then the
             database will be analysed at the appropriate places to
             ensure that database statistics are updated.
@@ -94,70 +123,72 @@ class Indexer:
         with connect(self.dsn) as conn:
             conn.autocommit = True
 
-            if analyse:
-                def _analyze():
+            def _analyze() -> None:
+                if analyse:
                     with conn.cursor() as cur:
                         cur.execute('ANALYZE')
-            else:
-                def _analyze():
-                    pass
 
-            self.index_by_rank(0, 4)
-            _analyze()
+            if self.index_by_rank(0, 4) > 0:
+                _analyze()
 
-            self.index_boundaries(0, 30)
-            _analyze()
+            if self.index_boundaries(0, 30) > 100:
+                _analyze()
 
-            self.index_by_rank(5, 25)
-            _analyze()
+            if self.index_by_rank(5, 25) > 100:
+                _analyze()
 
-            self.index_by_rank(26, 30)
-            _analyze()
+            if self.index_by_rank(26, 30) > 1000:
+                _analyze()
 
-            self.index_postcodes()
-            _analyze()
+            if self.index_postcodes() > 100:
+                _analyze()
 
 
-    def index_boundaries(self, minrank, maxrank):
+    def index_boundaries(self, minrank: int, maxrank: int) -> int:
         """ Index only administrative boundaries within the given rank range.
         """
+        total = 0
         LOG.warning("Starting indexing boundaries using %s threads",
                     self.num_threads)
 
-        for rank in range(max(minrank, 4), min(maxrank, 26)):
-            self._index(runners.BoundaryRunner(rank))
+        with self.tokenizer.name_analyzer() as analyzer:
+            for rank in range(max(minrank, 4), min(maxrank, 26)):
+                total += self._index(runners.BoundaryRunner(rank, analyzer))
+
+        return total
 
-    def index_by_rank(self, minrank, maxrank):
+    def index_by_rank(self, minrank: int, maxrank: int) -> int:
         """ Index all entries of placex in the given rank range (inclusive)
             in order of their address rank.
 
             When rank 30 is requested then also interpolations and
             places with address rank 0 will be indexed.
         """
+        total = 0
         maxrank = min(maxrank, 30)
         LOG.warning("Starting indexing rank (%i to %i) using %i threads",
                     minrank, maxrank, self.num_threads)
 
-        for rank in range(max(1, minrank), maxrank):
-            self._index(runners.RankRunner(rank))
+        with self.tokenizer.name_analyzer() as analyzer:
+            for rank in range(max(1, minrank), maxrank + 1):
+                total += self._index(runners.RankRunner(rank, analyzer), 20 if rank == 30 else 1)
+
+            if maxrank == 30:
+                total += self._index(runners.RankRunner(0, analyzer))
+                total += self._index(runners.InterpolationRunner(analyzer), 20)
 
-        if maxrank == 30:
-            self._index(runners.RankRunner(0))
-            self._index(runners.InterpolationRunner(), 20)
-            self._index(runners.RankRunner(30), 20)
-        else:
-            self._index(runners.RankRunner(maxrank))
+        return total
 
 
-    def index_postcodes(self):
-        """Index the entries ofthe location_postcode table.
+    def index_postcodes(self) -> int:
+        """Index the entries of the location_postcode table.
         """
         LOG.warning("Starting indexing postcodes using %s threads", self.num_threads)
 
-        self._index(runners.PostcodeRunner(), 20)
+        return self._index(runners.PostcodeRunner(), 20)
 
 
-    def update_status_table(self):
+    def update_status_table(self) -> None:
         """ Update the status in the status table to 'indexed'.
         """
         with connect(self.dsn) as conn:
@@ -166,7 +197,7 @@ class Indexer:
 
             conn.commit()
 
-    def _index(self, runner, batch=1):
+    def _index(self, runner: runners.Runner, batch: int = 1) -> int:
         """ Index a single rank or table. `runner` describes the SQL to use
             for indexing. `batch` describes the number of objects that
             should be processed with a single SQL statement
@@ -174,6 +205,7 @@ class Indexer:
         LOG.warning("Starting %s (using batch size %s)", runner.name(), batch)
 
         with connect(self.dsn) as conn:
+            psycopg2.extras.register_hstore(conn)
             with conn.cursor() as cur:
                 total_tuples = cur.scalar(runner.sql_count_objects())
                 LOG.debug("Total number of rows: %i", total_tuples)
@@ -186,20 +218,25 @@ class Indexer:
                 with conn.cursor(name='places') as cur:
                     cur.execute(runner.sql_get_objects())
 
-                    with WorkerPool(self.dsn, self.num_threads) as pool:
-                        while True:
-                            places = [p for p in cur.fetchmany(batch)]
-                            if not places:
-                                break
+                    with PlaceFetcher(self.dsn, conn) as fetcher:
+                        with WorkerPool(self.dsn, self.num_threads) as pool:
+                            has_more = fetcher.fetch_next_batch(cur, runner)
+                            while has_more:
+                                places = fetcher.get_batch()
 
-                            LOG.debug("Processing places: %s", str(places))
-                            worker = pool.next_free_worker()
+                                # asynchronously get the next batch
+                                has_more = fetcher.fetch_next_batch(cur, runner)
 
-                            runner.index_places(worker, places)
-                            progress.add(len(places))
+                                # And insert the current batch
+                                for idx in range(0, len(places), batch):
+                                    part = places[idx:idx + batch]
+                                    LOG.debug("Processing places: %s", str(part))
+                                    runner.index_places(pool.next_free_worker(), part)
+                                    progress.add(len(part))
 
-                        pool.finish_all()
+                            LOG.info("Wait time: fetcher: %.2fs,  pool: %.2fs",
+                                     fetcher.wait_time, pool.wait_time)
 
                 conn.commit()
 
-        progress.done()
+        return progress.done()