Multi-AZ and failover

[solracsf]

JuiceFS already supports one endpoint, let's say s3.us-east-1.amazonaws.com, by setting

--bucket bucket.s3.us-east-1.amazonaws.com

Problem here is that data will be stored in one region only us-east-1, so no Multi-AZ. Some providers allow bucket replication, but juiceFS wouldn't know how to reach them as only one endpoint is set...

My idea here was that if one could use 2 (or more) endpoints, data would be somehow replicated in these 2 buckets, in different regions, and JuiceFS could always try bucket 2 in case bucket 1 request fails for whatever reason for READ (not sure for WRITES) statements.

Imagine one can set this:

--bucket bucket1.s3.us-east-1.amazonaws.com
--failover bucket2.s3.us-east-2.amazonaws.com

As they both are accessed with the same access_key/secret_key, data would be somehow replicated over 2 regions.

WRITE request:

• Write to --bucket
• In the background, Write to --failover

Problem here is how to deal with fail requests; if either --bucket or --failover is not reachable, there would be 2 options:

• WRITE only in one of them, as READ would try --bucket first, followed by --failover, see below.
• WRITE in the available bucket, and replicate back to the other one in the background once it's available again (not sure if this could be done, and how).
  - Another approach?

READ requests are more "easy" to handle:

• Try --bucket
• If it fails (4xx or 5xx), try --failover

--failover option would be available for new filesystems only, not existing ones.

I don't know if this makes sense, but it could provide a good failover system and better availability and durability.

[davies]

@solracsf us-east-1 is a region with multiple available zones builtin, so any object written into us-east-1 is already replicated into multiple AZ by default, any failure of single AZ will not result in unavailability or data loss.

For replication across multiple regions or cloud vendors, is on the roadmap.