summary refs log tree commit diff
path: root/docs/server-server
diff options
context:
space:
mode:
authorKegan Dougal <kegan@matrix.org>2014-08-28 09:45:05 +0100
committerKegan Dougal <kegan@matrix.org>2014-08-28 09:45:05 +0100
commit660129deb1bb60dac04ff59d1da8bbe53882149e (patch)
tree79ca7c082e24e3039e75b3c0457cebd8a564f99a /docs/server-server
parentFix up the various presence-related tests so that if they're not skipped, the... (diff)
downloadsynapse-660129deb1bb60dac04ff59d1da8bbe53882149e.tar.xz
Shuffle files around in /docs
Diffstat (limited to 'docs/server-server')
-rw-r--r--docs/server-server/versioning.rst11
1 files changed, 11 insertions, 0 deletions
diff --git a/docs/server-server/versioning.rst b/docs/server-server/versioning.rst
new file mode 100644
index 0000000000..ffda60633f
--- /dev/null
+++ b/docs/server-server/versioning.rst
@@ -0,0 +1,11 @@
+Versioning is, like, hard for backfilling backwards because of the number of Home Servers involved.
+
+The way we solve this is by doing versioning as an acyclic directed graph of PDUs. For backfilling purposes, this is done on a per context basis. 
+When we send a PDU we include all PDUs that have been received for that context that hasn't been subsequently listed in a later PDU. The trivial case is a simple list of PDUs, e.g. A <- B <- C. However, if two servers send out a PDU at the same to, both B and C would point at A - a later PDU would then list both B and C.
+
+Problems with opaque version strings:
+    - How do you do clustering without mandating that a cluster can only have one transaction in flight to a given remote home server at a time. 
+      If you have multiple transactions sent at once, then you might drop one transaction, receive another with a version that is later than the dropped transaction and which point ARGH WE LOST A TRANSACTION.
+    - How do you do backfilling? A version string defines a point in a stream w.r.t. a single home server, not a point in the context.
+
+We only need to store the ends of the directed graph, we DO NOT need to do the whole one table of nodes and one of edges.