Difference between revisions of "NoDowntimeDBMigrations"
(→Online Schema Changes) |
(→Implementation Path) |
||
| Line 34: | Line 34: | ||
=== Implementation Path === | === Implementation Path === | ||
| − | Getting to the proposed solution will | + | Getting to the proposed solution will require non-trivial changes and will require some signifcant work. Smaller implementation steps should be taken to reduce the time it takes to get any improvements implemented. The goal is to incrementally reduce the amount of downtime with each step. |
==== Online Schema Changes ==== | ==== Online Schema Changes ==== | ||
Revision as of 17:10, 23 August 2013
Contents
Problem Description
In the past, some migrations have shown to take a significant amount of time run on large installations, causing unacceptable amounts of downtime during code upgrades.
As Openstack installations continue getting larger and larger, these long downtimes will continue getting worse and worse. If an Openstack provider guarantees 99.99% uptime, that leaves only 4 minutes total per month for all downtime (planned or unplanned). Some database migrations in the past have taken hours to run on large installations.
Schema Changes
MySQL will often grab a table lock during the entire take it takes to execute a schema change. This will block all reads and writes during the period the table lock is held. The length of time the lock is held often depends on the number of rows in the table.
MySQL will take a table lock depending on the version of the daemon, what underlying storage engine is used and what schema change is happening.
PostgreSQL is much better but will also grab a table lock in some cases (eg adding a column with a default value).
Schema and Data Format Expectations
Code in Openstack currently expects a fixed schema and data format. All schema changes and changes to data format are required to be completed after Openstack has been stopped and before the new version is running.
Proposed End Solution
Embrace the Expand/Contract pattern for database migrations. This pattern splits database migrations into three parts:
- Expand schema (adding new columns/tables/indexes)
- Migrate data
- Contract schema (removing unused columns/tables/indexes)
Code would migrate data on load when the service is running. Optionally, a background task can migrate all data at whatever speed specified. When all data that needs to be migrated has been migrated, the contraction can be run which will remove unused columns.
This decouples the database schema from the code allowing the two to be updated independently and allowing the service to continue running transparently while the data is migrated.
The goal is to get total downtime measured in seconds.
Implementation Path
Getting to the proposed solution will require non-trivial changes and will require some signifcant work. Smaller implementation steps should be taken to reduce the time it takes to get any improvements implemented. The goal is to incrementally reduce the amount of downtime with each step.
Online Schema Changes
In many cases, schema changes can be done while old software is still running. This is generally limited to cases where new tables, columns and indexes are added and where appropriate solutions that avoid causing reads/writes to block are available.
Some tools have been developed in the past to help with online schema changes:
Complicating any schema changes is the need to support all of the database Openstack currently supports: MySQL/MariaDB, PostgreSQL and SQLite. All of the existing solutions appear to be specific to MySQL/MariaDB. SQLite brings it's own set of problems by supporting basically no schema changes.
