Complement flake(?): GET_/rooms/:room_id/aliases_lists_aliases

[DMRobertson]

https://github.com/matrix-org/synapse/actions/runs/3249032216/jobs/5331241845

Edit(dmr): I am trying to reproduce this with

WORKERS=1 POSTGRES=1 COMPLEMENT_DIR=../complement scripts-dev/complement.sh -run 'TestRoomAlias/.*aliases_lists_aliases'

and have manually upped the replication torture setting in Synapse's source.

[DMRobertson]

https://github.com/matrix-org/synapse/actions/runs/3250081130/jobs/5333349627

[DMRobertson]

Why are we suddenly seeing this today?

https://github.com/matrix-org/synapse/actions/runs/3251269844/jobs/5336107983

[DMRobertson]

https://github.com/matrix-org/synapse/actions/runs/3252498613/jobs/5338825082

seems to be a worker-specific failure(?)

[DMRobertson]

https://github.com/matrix-org/synapse/actions/runs/3258600088/jobs/5350813055

[DMRobertson]

Test from matrix-org/complement#266.

Sytest still exists here: https://github.com/matrix-org/sytest/blame/da1c9e221f998b1d50214b9dd8ab18eb3804d57a/tests/10apidoc/32room-alias.pl#L52 and was deflaked matrix-org/sytest#1056. The deflaking was not ported to complement.

[DMRobertson]

Indeed, as Rich notes in matrix-org/sytest#1055 (comment):

the problem is that the request to set the alias now goes to the master, while the query goes to the client_reader.

[DMRobertson]

(I wonder if #14165 might have exposed this?)

[DMRobertson]

I've worked around this with a retry loop, but as noted here, the underlying problem is that Synapse (as an aggregation of workers) isn't offering read-after-write consistency here.

[DMRobertson]

https://github.com/matrix-org/synapse/actions/runs/3273203305/jobs/5385186367