Support for database schema version ranges

changelog.d/9933.misc

@@ -0,0 +1 @@
+Update the database schema versioning to support gradual migration away from legacy tables.

docs/SUMMARY.md

@@ -69,6 +69,7 @@
   - [Git Usage](dev/git.md)
   - [Testing]()
   - [OpenTracing](opentracing.md)
+  - [Database Schemas](development/database_schema.md)
   - [Synapse Architecture]()
     - [Log Contexts](log_contexts.md)
     - [Replication](replication.md)
@@ -84,4 +85,4 @@
   - [Scripts]()
 
 # Other
-  - [Dependency Deprecation Policy](deprecation_policy.md)
+  - [Dependency Deprecation Policy](deprecation_policy.md)

docs/development/database_schema.md

@@ -0,0 +1,95 @@
+# Synapse database schema files
+
+Synapse's database schema is stored in the `synapse.storage.schema` module.
+
+## Logical databases
+
+Synapse supports splitting its datastore across multiple physical databases (which can
+be useful for large installations), and the schema files are therefore split according
+to the logical database they are apply to.
+
+At the time of writing, the following "logical" databases are supported:
+
+* `state` - used to store Matrix room state (more specifically, `state_groups`,
+  their relationships and contents.)
+* `main` - stores everything else.
+
+Addionally, the `common` directory contains schema files for tables which must be
+present on *all* physical databases.
+
+## Synapse schema versions
+
+Synapse manages its database schema via "schema versions". These are mainly used to
+help avoid confusion if the Synapse codebase is rolled back after the database is
+updated. They work as follows:
+
+ * The Synapse codebase defines a constant `synapse.storage.schema.SCHEMA_VERSION`
+   which represents the expectations made about the database by that version. For
+   example, as of Synapse v1.36, this is `59`.
+
+ * The database stores a "compatibility version" in
+   `schema_compat_version.compat_version` which defines the `SCHEMA_VERSION` of the
+   oldest version of Synapse which will work with the database. On startup, if
+   `compat_version` is found to be newer than `SCHEMA_VERSION`, Synapse will refuse to
+   start.
+
+   Synapse automatically updates this field from
+   `synapse.storage.schema.SCHEMA_COMPAT_VERSION`.
+
+ * Whenever a backwards-incompatible change is made to the database format (normally
+   via a `delta` file), `synapse.storage.schema.SCHEMA_COMPAT_VERSION` is also updated
+   so that administrators can not accidentally roll back to a too-old version of Synapse.
+
+Generally, the goal is to maintain compatibility with at least one or two previous
+releases of Synapse, so any substantial change tends to require multiple releases and a
+bit of forward-planning to get right.
+
+As a worked example: we want to remove the `room_stats_historical` table. Here is how it
+might pan out.
+
+ 1. Replace any code that *reads* from `room_stats_historical` with alternative
+    implementations, but keep writing to it in case of rollback to an earlier version.
+    Also, increase `synapse.storage.schema.SCHEMA_VERSION`.  In this
+    instance, there is no existing code which reads from `room_stats_historical`, so
+    our starting point is:
+
+    v1.36.0: `SCHEMA_VERSION=59`, `SCHEMA_COMPAT_VERSION=59`
+
+ 2. Next (say in Synapse v1.37.0): remove the code that *writes* to
+    `room_stats_historical`, but don’t yet remove the table in case of rollback to
+    v1.36.0. Again, we increase `synapse.storage.schema.SCHEMA_VERSION`, but
+    because we have not broken compatibility with v1.36, we do not yet update
+    `SCHEMA_COMPAT_VERSION`. We now have:
+
+    v1.37.0: `SCHEMA_VERSION=60`, `SCHEMA_COMPAT_VERSION=59`.
+
+ 3. Later (say in Synapse v1.38.0): we can remove the table altogether. This will
+    break compatibility with v1.36.0, so we must update `SCHEMA_COMPAT_VERSION` accordingly.
+    There is no need to update `synapse.storage.schema.SCHEMA_VERSION`, since there is no
+    change to the Synapse codebase here. So we end up with:
+
+    v1.38.0: `SCHEMA_VERSION=60`, `SCHEMA_COMPAT_VERSION=60`.
+
+If in doubt about whether to update `SCHEMA_VERSION` or not, it is generally best to
+lean towards doing so.
+
+## Full schema dumps
+
+In the `full_schemas` directories, only the most recently-numbered snapshot is used
+(`54` at the time of writing). Older snapshots (eg, `16`) are present for historical
+reference only.
+
+### Building full schema dumps
+
+If you want to recreate these schemas, they need to be made from a database that
+has had all background updates run.
+
+To do so, use `scripts-dev/make_full_schema.sh`. This will produce new
+`full.sql.postgres` and `full.sql.sqlite` files.
+
+Ensure postgres is installed, then run:
+
+    ./scripts-dev/make_full_schema.sh -p postgres_username -o output_dir/
+
+NB at the time of writing, this script predates the split into separate `state`/`main`
+databases so will require updates to handle that correctly.

synapse/storage/prepare_database.py

@@ -1,5 +1,4 @@
-# Copyright 2014 - 2016 OpenMarket Ltd
-# Copyright 2018 New Vector Ltd
+# Copyright 2014 - 2021 The Matrix.org Foundation C.I.C.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -26,7 +25,7 @@
 from synapse.storage.database import LoggingDatabaseConnection
 from synapse.storage.engines import BaseDatabaseEngine
 from synapse.storage.engines.postgres import PostgresEngine
-from synapse.storage.schema import SCHEMA_VERSION
+from synapse.storage.schema import SCHEMA_COMPAT_VERSION, SCHEMA_VERSION
 from synapse.storage.types import Cursor
 
 logger = logging.getLogger(__name__)
@@ -59,6 +58,28 @@ class UpgradeDatabaseException(PrepareDatabaseException):
 )
 
 
+@attr.s
+class _SchemaState:
+    current_version: int = attr.ib()
+    """The current schema version of the database"""
+
+    compat_version: Optional[int] = attr.ib()
+    """The SCHEMA_VERSION of the oldest version of Synapse for this database
+
+    If this is None, we have an old version of the database without the necessary
+    table.
+    """
+
+    applied_deltas: Collection[str] = attr.ib(factory=tuple)
+    """Any delta files for `current_version` which have already been applied"""
+
+    upgraded: bool = attr.ib(default=False)
+    """Whether the current state was reached by applying deltas.
+
+    If False, we have run the full schema for `current_version`, and have applied no
+    deltas since. If True, we have run some deltas since the original creation."""
+
+
 def prepare_database(
     db_conn: LoggingDatabaseConnection,
     database_engine: BaseDatabaseEngine,
@@ -96,12 +117,11 @@ def prepare_database(
         version_info = _get_or_create_schema_state(cur, database_engine)
 
         if version_info:
-            user_version, delta_files, upgraded = version_info
             logger.info(
                 "%r: Existing schema is %i (+%i deltas)",
                 databases,
-                user_version,
-                len(delta_files),
+                version_info.current_version,
+                len(version_info.applied_deltas),
             )
 
             # config should only be None when we are preparing an in-memory SQLite db,
@@ -113,16 +133,18 @@ def prepare_database(
 
             # if it's a worker app, refuse to upgrade the database, to avoid multiple
             # workers doing it at once.
-            if config.worker_app is not None and user_version != SCHEMA_VERSION:
+            if (
+                config.worker_app is not None
+                and version_info.current_version != SCHEMA_VERSION
+            ):
                 raise UpgradeDatabaseException(
-                    OUTDATED_SCHEMA_ON_WORKER_ERROR % (SCHEMA_VERSION, user_version)
+                    OUTDATED_SCHEMA_ON_WORKER_ERROR
+                    % (SCHEMA_VERSION, version_info.current_version)
                 )
 
             _upgrade_existing_database(
                 cur,
-                user_version,
-                delta_files,
-                upgraded,
+                version_info,
                 database_engine,
                 config,
                 databases=databases,
@@ -261,9 +283,7 @@ def _setup_new_database(
 
     _upgrade_existing_database(
         cur,
-        current_version=max_current_ver,
-        applied_delta_files=[],
-        upgraded=False,
+        _SchemaState(current_version=max_current_ver, compat_version=None),
         database_engine=database_engine,
         config=None,
         databases=databases,
@@ -273,9 +293,7 @@ def _setup_new_database(
 
 def _upgrade_existing_database(
     cur: Cursor,
-    current_version: int,
-    applied_delta_files: List[str],
-    upgraded: bool,
+    current_schema_state: _SchemaState,
     database_engine: BaseDatabaseEngine,
     config: Optional[HomeServerConfig],
     databases: Collection[str],
@@ -321,12 +339,8 @@ def _upgrade_existing_database(
 
     Args:
         cur
-        current_version: The current version of the schema.
-        applied_delta_files: A list of deltas that have already been applied.
-        upgraded: Whether the current version was generated by having
-            applied deltas or from full schema file. If `True` the function
-            will never apply delta files for the given `current_version`, since
-            the current_version wasn't generated by applying those delta files.
+        current_schema_state: The current version of the schema, as
+            returned by _get_or_create_schema_state
         database_engine
         config:
             None if we are initialising a blank database, otherwise the application
@@ -337,13 +351,16 @@ def _upgrade_existing_database(
             upgrade portions of the delta scripts.
     """
     if is_empty:
-        assert not applied_delta_files
+        assert not current_schema_state.applied_deltas
     else:
         assert config
 
     is_worker = config and config.worker_app is not None
 
-    if current_version > SCHEMA_VERSION:
+    if (
+        current_schema_state.compat_version is not None
+        and current_schema_state.compat_version > SCHEMA_VERSION
+    ):
         raise ValueError(
             "Cannot use this database as it is too "
             + "new for the server to understand"
@@ -357,14 +374,26 @@ def _upgrade_existing_database(
         assert config is not None
         check_database_before_upgrade(cur, database_engine, config)
 
-    start_ver = current_version
+    # update schema_compat_version before we run any upgrades, so that if synapse
+    # gets downgraded again, it won't try to run against the upgraded database.
+    if (
+        current_schema_state.compat_version is None
+        or current_schema_state.compat_version < SCHEMA_COMPAT_VERSION
+    ):
+        cur.execute("DELETE FROM schema_compat_version")
+        cur.execute(
+            "INSERT INTO schema_compat_version(compat_version) VALUES (?)",
+            (SCHEMA_COMPAT_VERSION,),
+        )
+
+    start_ver = current_schema_state.current_version
 
     # if we got to this schema version by running a full_schema rather than a series
     # of deltas, we should not run the deltas for this version.
-    if not upgraded:
+    if not current_schema_state.upgraded:
         start_ver += 1
 
-    logger.debug("applied_delta_files: %s", applied_delta_files)
+    logger.debug("applied_delta_files: %s", current_schema_state.applied_deltas)
 
     if isinstance(database_engine, PostgresEngine):
         specific_engine_extension = ".postgres"
@@ -440,7 +469,7 @@ def _upgrade_existing_database(
             absolute_path = entry.absolute_path
 
             logger.debug("Found file: %s (%s)", relative_path, absolute_path)
-            if relative_path in applied_delta_files:
+            if relative_path in current_schema_state.applied_deltas:
                 continue
 
             root_name, ext = os.path.splitext(file_name)
@@ -621,25 +650,39 @@ def execute_statements_from_stream(cur: Cursor, f: TextIO) -> None:
 
 def _get_or_create_schema_state(
     txn: Cursor, database_engine: BaseDatabaseEngine
-) -> Optional[Tuple[int, List[str], bool]]:
+) -> Optional[_SchemaState]:
     # Bluntly try creating the schema_version tables.
     sql_path = os.path.join(schema_path, "common", "schema_version.sql")
     executescript(txn, sql_path)
 
     txn.execute("SELECT version, upgraded FROM schema_version")
     row = txn.fetchone()
 
+    if row is None:
+        # new database
+        return None
+
+    current_version = int(row[0])
+    upgraded = bool(row[1])
+
+    compat_version: Optional[int] = None
+    txn.execute("SELECT compat_version FROM schema_compat_version")
+    row = txn.fetchone()
     if row is not None:
-        current_version = int(row[0])
-        txn.execute(
-            "SELECT file FROM applied_schema_deltas WHERE version >= ?",
-            (current_version,),
-        )
-        applied_deltas = [d for d, in txn]
-        upgraded = bool(row[1])
-        return current_version, applied_deltas, upgraded
+        compat_version = int(row[0])
+
+    txn.execute(
+        "SELECT file FROM applied_schema_deltas WHERE version >= ?",
+        (current_version,),
+    )
+    applied_deltas = tuple(d for d, in txn)
 
-    return None
+    return _SchemaState(
+        current_version=current_version,
+        compat_version=compat_version,
+        applied_deltas=applied_deltas,
+        upgraded=upgraded,
+    )
 
 
 @attr.s(slots=True)

synapse/storage/schema/README.md

@@ -1,37 +1,4 @@
 # Synapse Database Schemas
 
-This directory contains the schema files used to build Synapse databases.
-
-Synapse supports splitting its datastore across multiple physical databases (which can
-be useful for large installations), and the schema files are therefore split according
-to the logical database they are apply to.
-
-At the time of writing, the following "logical" databases are supported:
-
-* `state` - used to store Matrix room state (more specifically, `state_groups`,
-  their relationships and contents.)
-* `main` - stores everything else.
-
-Addionally, the `common` directory contains schema files for tables which must be
-present on *all* physical databases.
-
-## Full schema dumps
-
-In the `full_schemas` directories, only the most recently-numbered snapshot is useful
-(`54` at the time of writing). Older snapshots (eg, `16`) are present for historical
-reference only.
-
-## Building full schema dumps
-
-If you want to recreate these schemas, they need to be made from a database that
-has had all background updates run.
-
-To do so, use `scripts-dev/make_full_schema.sh`. This will produce new
-`full.sql.postgres` and `full.sql.sqlite` files.
-
-Ensure postgres is installed, then run:
-
-    ./scripts-dev/make_full_schema.sh -p postgres_username -o output_dir/
-
-NB at the time of writing, this script predates the split into separate `state`/`main`
-databases so will require updates to handle that correctly.
+This directory contains the schema files used to build Synapse databases. For more
+information, see /docs/development/database_schema.md.