server: ensure SQL pods have unique instance IDs within the tenant

[knz]

In the serverless / multi-tenant architecture, we need to identify different SQL pods for the same tenant.

This separation is provided by the "Instance ID" abstraction in the source code. We use the instance ID at least in 3 places:

• when logging events, to distinguish servers under a single tenant.
• to compute the unique_rowid() function in SQL
• for table leases

As of this writing, the instance ID is not fully implemented and is fixed to a constant value. This prevents running multiple SQL pods for a single tenant.

We need to complete this by allocating a separate instance ID to each tenant process.

There are three design constraints:

• the value must remain 31 bits (like NodeID), and preferably use fewer bits than that.
• since we envision that SQL pods will be restarted very frequently, we can't blindly increment a sequence. Some form of ID reuse should be implemented.
• the instance ID must be known very early during server initialization, before the SQL subsystem is ready to serve queries. This means that we cannot look up the instance ID inside a SQL table from the pod process directly.

The third point alone suggests that the SQL pod process should request its SQL instance ID using a RPC to the nearest KV node.

Epic CRDB-2576

[knz]

cc @dhartunian

[exalate-issue-sync]

Assigned to rima