Motivation

Consider table's schema being altered just before the data transaction has started. Some nodes might have already applied the new table definition, while others not. As a result, when the transaction gets committed, it might get committed with different versions of the table schema on different nodes (or on some nodes the table might have been already considered dropped, while on others not), which breaks consistency.

To ensure consistency, we must use the same version of the table schema when committing a transaction across all participant nodes.

Additionally, it is imperative that a client is able to "view their own schema updates." In the event that a client initiates a schema update and receives confirmation that the update has been implemented, they must observe that the update has taken effect on any node they access.

Description

A schema update U for an object Q looks like "since the moment T_u, object Q has schema S". T_u is the activation moment of the new schema version S; in other words, S activates at T_u. Schema version S is effective from its activation moment T_u till the activation moment T_u’ of the next schema version S’ of the same object (including the start, excluding the end of the interval) or indefinitely if S is the most recent version of the schema at the moment; so it’s either [T_u, T_u’) or [T_u, +inf).

A schema update U is visible to a node N at moment T_op if the node N got a message M containing this update at a moment T<=T_op. (Schema update messages are distributed using Meta-Storage RAFT group (so a message consumption is an application of it to a state machine), so the update messages are delivered in order (hence, if a message is not delivered, none of the messages that follow it will be delivered) and not duplicated)). This is due to the fact that there is exactly one Meta-Storage RAFT state machine at each node.

We want that for any given timestamp T_q, all nodes get the same answer to the question: ‘whether a schema update U is visible and effective or not at the moment T_q’.

This can be reformulated: a schema update saying that a new schema version activates at T_u splits all transactions into two sets: those that commit strictly before T_u (T_c < T_u) and hence operate on the old schema version, and those that commit starting with T_u (T_c ≥ T_u) and hence operate on the new schema version, regardless of the nodes that participate in the corresponding transaction.

An important thing here is the clock we use to define the timestamps. 

As the nodes’ clocks do not have to be synchronized, it’s the node’s local resolution of "whether a timestamp is reached by the node’s clock or not" that matters, it’s not relative to some absolute clock (about which each node only has an approximate idea).

Let’s call the property defined above the Atomicity Property (AP).

The following property makes AP to hold: no node can get a message about an update U(T_u,S) later than the node has fetched (and used in handling some transaction) a schema for a moment T_op >= T_u.

To maintain the real-time atomicity, we would need to coordinate between all nodes of the cluster on each schema update to make sure they get the update before they can commit a transaction using it, so a lagging node would make all nodes wait (not processing any transactions); alternatively, if a schema update might not be acknowledged by all nodes in a timely fashion, a schema update could be held in an inactive state for an indeterminate period of time creating a form of unavailability of the whole cluster.

That's why we separate the timestamps assigned to the events (like schema version activation and schema version fetches) from the actual moments when the corresponding events tagged with these timestamps are processed on each node. We have:

1. A monotonic clock from which we take timestamps used to tag events. The timestamps are totally ordered. These are reference timestamps on a reference timeline.
2. Events tagged with these timestamps (hence, also totally ordered)
3. On each node, these events are ‘executed’ or ‘applied’ in the node timeline in the same total order, but this execution/application might happen with some lag after the events’ timestamps were taken from the clock. Distances between application moments of the corresponding events do not have to be equal (or proportional to) the distances between the corresponding timestamps labeling the events: only the ordering matters.
4. In real time, the application of events might happen in different moments of time on different nodes.

Schema change flow

1. Client issues a statement that modifies database schema (CREATE/ALTER/DROP TABLE, CREATE/DROP INDEX, etc)
2. The new schema version (with its T_u) is propagated to all cluster nodes using the Meta-storage (as we have a Meta-Storage RAFT group member, a learner or not, on every cluster node)
3. Each node acts independently from other nodes in interpreting the T_u, so no cluster-wide coordination between nodes is needed when partitioning the transactions. Schema version activation happens "automatically" when the node is sure it has all the updates it needs for processing the transaction it is processing. To make sure it has all the updates, a node might need to wait.
4. Before an update is reported as installed to the client, an additional wait happens to make sure that the client’s request on any node will see the installed update after the installation call returns. The required wait is until the schema update activation becomes non-future on all nodes (taking clock skew into account, so the client must wait for at least Now ≥ T_u+CS_max).

A simplified diagram demonstrating the flow of information on a schema update (low-level details (timestamp arithmetics for optimization) are not shown):

Risks and Assumptions

Schema updates are not visible right away

Even on the Meta-storage leader, a just-installed schema update is not visible right away, because it activates after DD in future. This might be confusing for a user that just issues an ALTER TABLE query because ‘read own writes’ is violated. This can be alleviated by making sure that we complete the future used to install the schema update not later than DD passes, but if a client fails before the wait elapses and reconnects, they might expect to see the update.

A node not receiving Meta-storage updates freezes processing of transactions where the node participate

A primary replica needs to wait for safeTime to reach T_op-DD before processing a transaction at a reference timestamp T_op. If the node does not get any Meta-Storage updates (i.e. it lags behind the Meta-Storage leader), it cannot proceed.

Possible situations:

1. The primary replica actually lags behind the whole cluster. In such a situation, we need to promote another replica to become a primary. This case is not caused by the new mechanism introduced here, we already should have measures to fight this.
2. The whole cluster lags behind the Meta-storage leader. This means that the whole cluster is in an uncontrollable state. In such a situation, we become unable to process transactions.

Note that if a primary P lags, then transactions where P does not participate are not affected.

An unpleasant thing is that a node lagging too much behind the Meta-Storage freezes transaction processing even in the absence of any schema updates.

Discussion Links

// N/A

Reference Links

// N/A

Tickets

// Links or report with relevant JIRA tickets.