Add MultiWriterIdGenerator. (#7281)

This will be used to coordinate stream IDs across multiple writers.

Functions as the equivalent of both `StreamIdGenerator` and
`SlavedIdTracker`.

1 files changed, 167 insertions, 2 deletions

diff --git a/synapse/storage/util/id_generators.py b/synapse/storage/util/id_generators.py
index 9d851beaa5..86d04ea9ac 100644
--- a/synapse/storage/util/id_generators.py
+++ b/synapse/storage/util/id_generators.py
@@ -16,6 +16,11 @@
 import contextlib
 import threading
 from collections import deque
+from typing import Dict, Set, Tuple
+
+from typing_extensions import Deque
+
+from synapse.storage.database import Database, LoggingTransaction
 
 
 class IdGenerator(object):
@@ -87,7 +92,7 @@ class StreamIdGenerator(object):
             self._current = (max if step > 0 else min)(
                 self._current, _load_current_id(db_conn, table, column, step)
             )
-        self._unfinished_ids = deque()
+        self._unfinished_ids = deque()  # type: Deque[int]
 
     def get_next(self):
         """
@@ -163,7 +168,7 @@ class ChainedIdGenerator(object):
         self.chained_generator = chained_generator
         self._lock = threading.Lock()
         self._current_max = _load_current_id(db_conn, table, column)
-        self._unfinished_ids = deque()
+        self._unfinished_ids = deque()  # type: Deque[Tuple[int, int]]
 
     def get_next(self):
         """
@@ -198,3 +203,163 @@ class ChainedIdGenerator(object):
                 return stream_id - 1, chained_id
 
             return self._current_max, self.chained_generator.get_current_token()
+
+
+class MultiWriterIdGenerator:
+    """An ID generator that tracks a stream that can have multiple writers.
+
+    Uses a Postgres sequence to coordinate ID assignment, but positions of other
+    writers will only get updated when `advance` is called (by replication).
+
+    Note: Only works with Postgres.
+
+    Args:
+        db_conn
+        db
+        instance_name: The name of this instance.
+        table: Database table associated with stream.
+        instance_column: Column that stores the row's writer's instance name
+        id_column: Column that stores the stream ID.
+        sequence_name: The name of the postgres sequence used to generate new
+            IDs.
+    """
+
+    def __init__(
+        self,
+        db_conn,
+        db: Database,
+        instance_name: str,
+        table: str,
+        instance_column: str,
+        id_column: str,
+        sequence_name: str,
+    ):
+        self._db = db
+        self._instance_name = instance_name
+        self._sequence_name = sequence_name
+
+        # We lock as some functions may be called from DB threads.
+        self._lock = threading.Lock()
+
+        self._current_positions = self._load_current_ids(
+            db_conn, table, instance_column, id_column
+        )
+
+        # Set of local IDs that we're still processing. The current position
+        # should be less than the minimum of this set (if not empty).
+        self._unfinished_ids = set()  # type: Set[int]
+
+    def _load_current_ids(
+        self, db_conn, table: str, instance_column: str, id_column: str
+    ) -> Dict[str, int]:
+        sql = """
+            SELECT %(instance)s, MAX(%(id)s) FROM %(table)s
+            GROUP BY %(instance)s
+        """ % {
+            "instance": instance_column,
+            "id": id_column,
+            "table": table,
+        }
+
+        cur = db_conn.cursor()
+        cur.execute(sql)
+
+        # `cur` is an iterable over returned rows, which are 2-tuples.
+        current_positions = dict(cur)
+
+        cur.close()
+
+        return current_positions
+
+    def _load_next_id_txn(self, txn):
+        txn.execute("SELECT nextval(?)", (self._sequence_name,))
+        (next_id,) = txn.fetchone()
+        return next_id
+
+    async def get_next(self):
+        """
+        Usage:
+            with await stream_id_gen.get_next() as stream_id:
+                # ... persist event ...
+        """
+        next_id = await self._db.runInteraction("_load_next_id", self._load_next_id_txn)
+
+        # Assert the fetched ID is actually greater than what we currently
+        # believe the ID to be. If not, then the sequence and table have got
+        # out of sync somehow.
+        assert self.get_current_token() < next_id
+
+        with self._lock:
+            self._unfinished_ids.add(next_id)
+
+        @contextlib.contextmanager
+        def manager():
+            try:
+                yield next_id
+            finally:
+                self._mark_id_as_finished(next_id)
+
+        return manager()
+
+    def get_next_txn(self, txn: LoggingTransaction):
+        """
+        Usage:
+
+            stream_id = stream_id_gen.get_next(txn)
+            # ... persist event ...
+        """
+
+        next_id = self._load_next_id_txn(txn)
+
+        with self._lock:
+            self._unfinished_ids.add(next_id)
+
+        txn.call_after(self._mark_id_as_finished, next_id)
+        txn.call_on_exception(self._mark_id_as_finished, next_id)
+
+        return next_id
+
+    def _mark_id_as_finished(self, next_id: int):
+        """The ID has finished being processed so we should advance the
+        current poistion if possible.
+        """
+
+        with self._lock:
+            self._unfinished_ids.discard(next_id)
+
+            # Figure out if its safe to advance the position by checking there
+            # aren't any lower allocated IDs that are yet to finish.
+            if all(c > next_id for c in self._unfinished_ids):
+                curr = self._current_positions.get(self._instance_name, 0)
+                self._current_positions[self._instance_name] = max(curr, next_id)
+
+    def get_current_token(self, instance_name: str = None) -> int:
+        """Gets the current position of a named writer (defaults to current
+        instance).
+
+        Returns 0 if we don't have a position for the named writer (likely due
+        to it being a new writer).
+        """
+
+        if instance_name is None:
+            instance_name = self._instance_name
+
+        with self._lock:
+            return self._current_positions.get(instance_name, 0)
+
+    def get_positions(self) -> Dict[str, int]:
+        """Get a copy of the current positon map.
+        """
+
+        with self._lock:
+            return dict(self._current_positions)
+
+    def advance(self, instance_name: str, new_id: int):
+        """Advance the postion of the named writer to the given ID, if greater
+        than existing entry.
+        """
+
+        with self._lock:
+            self._current_positions[instance_name] = max(
+                new_id, self._current_positions.get(instance_name, 0)
+            )