Arvados

h1. Multi-cluster user database

It is sometimes desirable to share a single user database across multiple Arvados clusters. For example:

* Clusters aaaaa, bbbbb, ccccc are on different continents.

* A down/unreachable cluster should not prevent a user from accessing _other_ clusters -- even if the down/unreachable cluster is normally the best/default one from that user's perspective.

This requires some changes to authentication (obtaining and validating API tokens).

h2. Obtaining tokens

Each user must be able to log in to their account using any cluster, regardless of where/whether they have logged in previously. This contrasts with the current setup, where each user account has a "home cluster" which must be used to log in.

To achieve this (without depending real-time communication between clusters) we need all of the participating clusters to agree on a mapping of upstream authentication results to Arvados user UUIDs. For example, if the upstream authentication result is @"ldap://ldap.example foo@bar.example"@ ("ldap://ldap.example assures us this user is foo@bar.example"):

# If a row already exists in the users table with <code>upstream == "ldap://ldap.example foo@bar.example"</code> then use that row

# Otherwise, create a new row with user UUID "fffff-tpzed-${sha1part(upstream)}" (where fffff is a common prefix used by all participating clusters and sha1part() is the first 15 chars of base-36-encoded sha1())

To avoid changing existing user accounts' UUIDs after this change, we would do a one-time synchronization across all participating clusters. For example, if "aaaaa-tpzed-012340123401234" exists on cluster aaaaa, we would add that row to bbbbb and ccccc as well. Next time that user logs in to bbbbb, bbbbb would issue a token itself, rather than deferring to aaaaa.

h2. Validating tokens

Each cluster must be able to validate a token that was issued by a different, currently unreachable, cluster. This contrasts with the current setup, where aaaaa validates tokens issued by bbbbb by doing a callback to bbbbb.

This seems easy enough: instead of random strings, tokens can be [like] JWT, signed by a private key whose public part is known by all clusters. (This would be more efficient than callbacks even for mutually untrusted clusters.)