Core Platform Team/Initiatives/Serve read requests from multiple datacenters
Serve read requests from multiple datacenters
Currently we are serving all of our MediaWiki-related traffic from a single DC (data centre), which is problematic for a number of reasons:
- if the main DC gets cut off or experiences networking issues, all of our projects become invisible to users;
- we keep a full copy of the infrastructure in the secondary DC, thus wasting hardware and electricity; and
- serving content from only one DC means slower access times and higher latencies for markets that would profit from being served from other DCs
As a first step in being able to fully serve traffic from multiple DCs, we need to start serving only GET requests (a.k.a. read requests) from multiple DCs.
- Significance and Motivation
By completing this project, we will have better utilisation of the hardware we have in all DCs and partially avoid catastrophic events. Moreover, in doing so we will be able to serve better and faster emergent markets.
Increase the scalability of the platform for future applications and new types of content, as well as a growing user base and amount of content
The primary measure of the success of this project is that both MediaWiki and the production environment can serve GET requests in all DCs.
- Baseline Metrics
- Infrastructure in CODFW ready to serve read requests without slowing down all writes and reads in all DCs
- Target Metrics
- Split incoming traffic on the edge for GET requests and route them to the closest DC (instead of the main DC)
- Core Platform
- Known Dependencies/Blockers
Epics, User Stories, and Requirements
- Replace all the DB slaves in codfw (need new hardware to support traffic)
- Update and/or replace GTID_Wait / heartbeat
- Decide where to move mainstash data (non-session data in Redis)
- Migrate mainstash data to new location
- Decide what is the acceptable replication lag for MediaWiki?
- Current lag is 10-15 seconds. ~1 second would be acceptable for 90% of reads get what they need, last 10% needs reengineering
- Evaluate lag for codfw, if not acceptable, then engineer a solution
- Update MediaWiki code to wait for replication (if needed)
- TLS Proxy work
- Separate traffic on varnish level
- Serve thumbnails
Time and Resource Estimates
- Estimated Start Date
- Actual Start Date
Started July 2018
- Estimated Completion Date
- Actual Completion Date
- Resource Estimates
- Core Platform
Many parts of MediaWiki assume that the DB is "close by and secure", which will change for Multi-DC. How do we address this?
Long term, what do we do with MYSQL? It doesn’t do master-master operations and blocks us from supporting writes in both locations.
If we remove performance tricks in MediaWiki, would the performance be acceptable and enable us to generalize the DB abstraction? Working on the Abstract Schema Migrations RFC may clear some of this question up.