Add a new Redis instance

GitLab can make use of multiple Redis instances. These instances are functionally partitioned so that, for example, we can store CI trace chunks from one Redis instance while storing sessions in another.

From time to time we might want to add a new Redis instance. Typically this will be a functional partition split from one of the existing instances such as the cache or shared state. This document describes an approach for adding a new Redis instance that handles existing data, based on prior examples:

• Dedicated Redis instance for Trace Chunk storage.
• Create dedicated Redis instance for Rate Limiting data.

This document does not cover the operational side of preparing and configuring the new Redis instance in detail, but the example epics do contain information on previous approaches to this.

Step 1: Support configuring the new instance

Before we can switch any features to using the new instance, we have to support configuring it and referring to it in the codebase. We must support the main installation types:

• Self-compiled installations (including development environments) - example MR
• Linux package installations - example MR
• Helm charts - example MR

Fallback instance

In the application code, we need to define a fallback instance in case the new instance is not configured. For example, if a GitLab instance has already configured a separate shared state Redis, and we are partitioning data from the shared state Redis, our new instance's configuration should default to that of the shared state Redis when it's not present. Otherwise we could break instances that don't configure the new Redis instance as soon as it's available.

You can define a .config_fallback method in Gitlab::Redis::Wrapper (the base class for all Redis instances) that defines the instance to be used if this one is not configured. If we were adding a Foo instance that should fall back to SharedState, we can do that like this:

module Gitlab
  module Redis
    class Foo < ::Gitlab::Redis::Wrapper
      # The data we store on Foo used to be stored on SharedState.
      def self.config_fallback
        SharedState
      end
    end
  end
end

We should also add specs like those in trace_chunks_spec.rb to ensure that this fallback works correctly.

Step 2: Support writing to and reading from the new instance

When migrating to the new instance, we must account for cases where data is either on:

• The 'old' (original) instance.
• The new one that we have just added support for.

As a result we may need to support reading from and writing to both instances, depending on some condition.

The exact condition to use varies depending on the data to be migrated. For the trace chunks case above, there was already a database column indicating where the data was stored (as there are other storage options than Redis).

This step may not apply if the data has a very short lifetime (a few minutes at most) and is not critical. In that case, we may decide that it is OK to incur a small amount of data loss and switch over through configuration only.

If there is not a more natural way to mark where the data is stored, using a feature flag may be convenient:

• It does not require an application restart to take effect.
• It applies to all application instances (Sidekiq, API, web, etc.) at the same time.
• It supports incremental rollout - ideally by actor (project, group, user, etc.) - so that we can monitor for errors and roll back easily.

Step 3: Migrate the data

We then need to configure the new instance for GitLab.com's production and staging environments. Hopefully it will be possible to test this change effectively on staging, to at least make sure that basic usage continues to work.

After that is done, we can roll out the change to production. Ideally this would be in an incremental fashion, following the standard incremental rollout documentation for feature flags.

When we have been using the new instance 100% of the time in production for a while and there are no issues, we can proceed.

Proposed solution: Migrate data by using MultiStore with the fallback strategy

We need a way to migrate users to a new Redis store without causing any inconveniences from UX perspective. We also want the ability to fall back to the "old" Redis instance if something goes wrong with the new instance.

Migration Requirements:

• No downtime.
• No loss of stored data until the TTL for storing data expires.
• Partial rollout using feature flags or ENV vars or combinations of both.
• Monitoring of the switch.
• Prometheus metrics in place.
• Easy rollback without downtime in case the new instance or logic does not behave as expected.

It is somewhat similar to the zero-downtime DB table rename. We need to write data into both Redis instances (old + new). We read from the new instance, but we need to fall back to the old instance when pre-fetching from the new dedicated Redis instance that failed. We need to log any issues or exceptions with a new instance, but still fall back to the old instance.

The proposed migration strategy is to implement and use the MultiStore. We used this approach with adding new dedicated Redis instance for session keys. Also MultiStore comes with corresponding specs.

The MultiStore looks like a redis-rb ::Redis instance.

In the new Redis instance class you added in Step 1, override the Redis method from the ::Gitlab::Redis::Wrapper

module Gitlab
  module Redis
    class Foo < ::Gitlab::Redis::Wrapper
      ...
      def self.redis
        # Don't use multistore if redis.foo configuration is not provided
        return super if config_fallback?

        primary_store = ::Redis.new(params)
        secondary_store = ::Redis.new(config_fallback.params)

        MultiStore.new(primary_store, secondary_store, store_name)
      end
    end
  end
end

MultiStore is initialized by providing the new Redis instance as a primary store, and old (fallback-instance) as a secondary store. The third argument is store_name which is used for logs, metrics and feature flag names, in case we use MultiStore implementation for different Redis stores at the same time.

By default, the MultiStore reads and writes only from the default Redis store. The default Redis store is secondary_store (the old fallback-instance). This allows us to introduce MultiStore without changing the default behavior.

MultiStore uses two feature flags to control the actual migration:

• use_primary_and_secondary_stores_for_[store_name]
• use_primary_store_as_default_for_[store_name]

For example, if our new Redis instance is called Gitlab::Redis::Foo, we can create two feature flags by executing:

bin/feature-flag use_primary_and_secondary_stores_for_foo
bin/feature-flag use_primary_store_as_default_for_foo

By enabling use_primary_and_secondary_stores_for_foo feature flag, our Gitlab::Redis::Foo will use MultiStore to write to both new Redis instance and the old (fallback-instance). All read commands are performed only on the default store which is controlled using the use_primary_store_as_default_for_foo feature flag. By enabling use_primary_store_as_default_for_foo feature flag, the MultiStore uses primary_store (new instance) as default Redis store.

For pipelined commands (pipelined and multi), we execute the entire operation in both stores and then compare the results. If they differ, we emit a Gitlab::Redis::MultiStore:PipelinedDiffError error, and track it in the gitlab_redis_multi_store_pipelined_diff_error_total Prometheus counter.

After a period of time for the new store to be populated, we can perform external validation to compare the state of both stores. Upon satisfactory validation results, we are probably safe to move the traffic to the new Redis store. We can disable use_primary_and_secondary_stores_for_foo feature flag. This will allow the MultiStore to read and write only from the primary Redis store (new store), moving all the traffic to the new Redis store.

Once we have moved all our traffic to the primary store, our data migration is complete. We can safely remove the MultiStore implementation and continue to use newly introduced Redis store instance.

Implementation details

MultiStore implements read and write Redis commands separately.

Read commands

• get

• mget

• smembers

• scard

• exists

• exists?

• get

• hexists

• hget

• hgetall

• hlen

• hmget

• hscan_each

• mapped_hmget

• mget

• scan_each

• scard

• sismember

• smembers

• sscan

• sscan_each

• ttl

• zscan_each

Write commands

• del
• eval
• expire
• flushdb
• hdel
• hset
• incr
• incrby
• mapped_hmset
• rpush
• sadd
• set
• setex
• setnx
• srem
• unlink

pipelined commands

NOTE: The Ruby block passed to these commands will be executed twice, once per each store. Thus, excluding the Redis operations performed, the block should be idempotent.

• pipelined
• multi

When a command outside of the supported list is used, method_missing will pass it to the old Redis instance and keep track of it. This ensures that anything unexpected behaves like it would before. In development or test environment, an error would be raised for early detection.

NOTE: By tracking gitlab_redis_multi_store_method_missing_total counter and Gitlab::Redis::MultiStore::MethodMissingError, a developer will need to add an implementation for missing Redis commands before proceeding with the migration.

Errors

Metrics

Step 4: clean up after the migration

We may choose to keep the migration paths or remove them, depending on whether or not we expect self-managed instances to perform this migration. gitlab-com/gl-infra/scalability#1131 contains a discussion on this topic for the trace chunks feature flag. It may be - as in that case - that we decide that the maintenance costs of supporting the migration code are higher than the benefits of allowing self-managed instances to perform this migration seamlessly, if we expect self-managed instances to cope without this functional partition.

If we decide to keep the migration code:

• We should document the migration steps.
• If we used a feature flag, we should ensure it's an ops type feature flag, as these are long-lived flags.

Otherwise, we can remove the flags and conclude the project.