summary refs log tree commit diff
path: root/synapse/storage/client_ips.py
diff options
context:
space:
mode:
authorRichard van der Hoff <richard@matrix.org>2019-02-14 14:41:40 +0000
committerRichard van der Hoff <richard@matrix.org>2019-02-14 14:41:40 +0000
commit00cf679bf22737b6b9028ba0dada47e029fc0925 (patch)
tree329a57fbbde7507c8715f8fde3f2669bc21beea2 /synapse/storage/client_ips.py
parentUpdate MSC1711_certificates_FAQ.md (diff)
parent0.99.1 (diff)
downloadsynapse-00cf679bf22737b6b9028ba0dada47e029fc0925.tar.xz
Merge tag 'v0.99.1'
Synapse 0.99.1 (2019-02-14)
===========================

Features
--------

- Include m.room.encryption on invites by default ([\#3902](https://github.com/matrix-org/synapse/issues/3902))
- Federation OpenID listener resource can now be activated even if federation is disabled ([\#4420](https://github.com/matrix-org/synapse/issues/4420))
- Synapse's ACME support will now correctly reprovision a certificate that approaches its expiry while Synapse is running. ([\#4522](https://github.com/matrix-org/synapse/issues/4522))
- Add ability to update backup versions ([\#4580](https://github.com/matrix-org/synapse/issues/4580))
- Allow the "unavailable" presence status for /sync.
  This change makes Synapse compliant with r0.4.0 of the Client-Server specification. ([\#4592](https://github.com/matrix-org/synapse/issues/4592))
- There is no longer any need to specify `no_tls`: it is inferred from the absence of TLS listeners ([\#4613](https://github.com/matrix-org/synapse/issues/4613), [\#4615](https://github.com/matrix-org/synapse/issues/4615), [\#4617](https://github.com/matrix-org/synapse/issues/4617), [\#4636](https://github.com/matrix-org/synapse/issues/4636))
- The default configuration no longer requires TLS certificates. ([\#4614](https://github.com/matrix-org/synapse/issues/4614))

Bugfixes
--------

- Copy over room federation ability on room upgrade. ([\#4530](https://github.com/matrix-org/synapse/issues/4530))
- Fix noisy "twisted.internet.task.TaskStopped" errors in logs ([\#4546](https://github.com/matrix-org/synapse/issues/4546))
- Synapse is now tolerant of the `tls_fingerprints` option being None or not specified. ([\#4589](https://github.com/matrix-org/synapse/issues/4589))
- Fix 'no unique or exclusion constraint' error ([\#4591](https://github.com/matrix-org/synapse/issues/4591))
- Transfer Server ACLs on room upgrade. ([\#4608](https://github.com/matrix-org/synapse/issues/4608))
- Fix failure to start when not TLS certificate was given even if TLS was disabled. ([\#4618](https://github.com/matrix-org/synapse/issues/4618))
- Fix self-signed cert notice from generate-config. ([\#4625](https://github.com/matrix-org/synapse/issues/4625))
- Fix performance of `user_ips` table deduplication background update ([\#4626](https://github.com/matrix-org/synapse/issues/4626), [\#4627](https://github.com/matrix-org/synapse/issues/4627))

Internal Changes
----------------

- Change the user directory state query to use a filtered call to the db instead of a generic one. ([\#4462](https://github.com/matrix-org/synapse/issues/4462))
- Reject federation transactions if they include more than 50 PDUs or 100 EDUs. ([\#4513](https://github.com/matrix-org/synapse/issues/4513))
- Reduce duplication of ``synapse.app`` code. ([\#4567](https://github.com/matrix-org/synapse/issues/4567))
- Fix docker upload job to push -py2 images. ([\#4576](https://github.com/matrix-org/synapse/issues/4576))
- Add port configuration information to ACME instructions. ([\#4578](https://github.com/matrix-org/synapse/issues/4578))
- Update MSC1711 FAQ to calrify .well-known usage ([\#4584](https://github.com/matrix-org/synapse/issues/4584))
- Clean up default listener configuration ([\#4586](https://github.com/matrix-org/synapse/issues/4586))
- Clarifications for reverse proxy docs ([\#4607](https://github.com/matrix-org/synapse/issues/4607))
- Move ClientTLSOptionsFactory init out of `refresh_certificates` ([\#4611](https://github.com/matrix-org/synapse/issues/4611))
- Fail cleanly if listener config lacks a 'port' ([\#4616](https://github.com/matrix-org/synapse/issues/4616))
- Remove redundant entries from docker config ([\#4619](https://github.com/matrix-org/synapse/issues/4619))
- README updates ([\#4621](https://github.com/matrix-org/synapse/issues/4621))
Diffstat (limited to 'synapse/storage/client_ips.py')
-rw-r--r--synapse/storage/client_ips.py87
1 files changed, 84 insertions, 3 deletions
diff --git a/synapse/storage/client_ips.py b/synapse/storage/client_ips.py

index 091d7116c5..9c21362226 100644 --- a/synapse/storage/client_ips.py +++ b/synapse/storage/client_ips.py
@@ -66,6 +66,11 @@ class ClientIpStore(background_updates.BackgroundUpdateStore): ) self.register_background_update_handler( + "user_ips_analyze", + self._analyze_user_ip, + ) + + self.register_background_update_handler( "user_ips_remove_dupes", self._remove_user_ip_dupes, ) @@ -109,6 +114,25 @@ class ClientIpStore(background_updates.BackgroundUpdateStore): defer.returnValue(1) @defer.inlineCallbacks + def _analyze_user_ip(self, progress, batch_size): + # Background update to analyze user_ips table before we run the + # deduplication background update. The table may not have been analyzed + # for ages due to the table locks. + # + # This will lock out the naive upserts to user_ips while it happens, but + # the analyze should be quick (28GB table takes ~10s) + def user_ips_analyze(txn): + txn.execute("ANALYZE user_ips") + + yield self.runInteraction( + "user_ips_analyze", user_ips_analyze + ) + + yield self._end_background_update("user_ips_analyze") + + defer.returnValue(1) + + @defer.inlineCallbacks def _remove_user_ip_dupes(self, progress, batch_size): # This works function works by scanning the user_ips table in batches # based on `last_seen`. For each row in a batch it searches the rest of @@ -167,12 +191,16 @@ class ClientIpStore(background_updates.BackgroundUpdateStore): clause = "? <= last_seen AND last_seen < ?" args = (begin_last_seen, end_last_seen) + # (Note: The DISTINCT in the inner query is important to ensure that + # the COUNT(*) is accurate, otherwise double counting may happen due + # to the join effectively being a cross product) txn.execute( """ SELECT user_id, access_token, ip, - MAX(device_id), MAX(user_agent), MAX(last_seen) + MAX(device_id), MAX(user_agent), MAX(last_seen), + COUNT(*) FROM ( - SELECT user_id, access_token, ip + SELECT DISTINCT user_id, access_token, ip FROM user_ips WHERE {} ) c @@ -186,7 +214,60 @@ class ClientIpStore(background_updates.BackgroundUpdateStore): # We've got some duplicates for i in res: - user_id, access_token, ip, device_id, user_agent, last_seen = i + user_id, access_token, ip, device_id, user_agent, last_seen, count = i + + # We want to delete the duplicates so we end up with only a + # single row. + # + # The naive way of doing this would be just to delete all rows + # and reinsert a constructed row. However, if there are a lot of + # duplicate rows this can cause the table to grow a lot, which + # can be problematic in two ways: + # 1. If user_ips is already large then this can cause the + # table to rapidly grow, potentially filling the disk. + # 2. Reinserting a lot of rows can confuse the table + # statistics for postgres, causing it to not use the + # correct indices for the query above, resulting in a full + # table scan. This is incredibly slow for large tables and + # can kill database performance. (This seems to mainly + # happen for the last query where the clause is simply `? < + # last_seen`) + # + # So instead we want to delete all but *one* of the duplicate + # rows. That is hard to do reliably, so we cheat and do a two + # step process: + # 1. Delete all rows with a last_seen strictly less than the + # max last_seen. This hopefully results in deleting all but + # one row the majority of the time, but there may be + # duplicate last_seen + # 2. If multiple rows remain, we fall back to the naive method + # and simply delete all rows and reinsert. + # + # Note that this relies on no new duplicate rows being inserted, + # but if that is happening then this entire process is futile + # anyway. + + # Do step 1: + + txn.execute( + """ + DELETE FROM user_ips + WHERE user_id = ? AND access_token = ? AND ip = ? AND last_seen < ? + """, + (user_id, access_token, ip, last_seen) + ) + if txn.rowcount == count - 1: + # We deleted all but one of the duplicate rows, i.e. there + # is exactly one remaining and so there is nothing left to + # do. + continue + elif txn.rowcount >= count: + raise Exception( + "We deleted more duplicate rows from 'user_ips' than expected", + ) + + # The previous step didn't delete enough rows, so we fallback to + # step 2: # Drop all the duplicates txn.execute(