...

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.

...

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 labelling the events: only the ordering matters.
4. In real time, the application of events might happen in different moments of time on different nodes.

...

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 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):

Making sure a node does not miss any schema messages that are relevant to a reference timestamp

The clocks may be skewed and the network is asynchronous (so message delivery might take an arbitrarily long time). This means that we cannot just use the naive approach of installing the update to the future and processing transactions on the nodes right away.

Briefly, the idea is to do the following:

1. Use the fact that Meta Storage is available on all cluster nodes (thanks to Learners) to distribute the schema updates
2. Wait for Meta Storage safeTime on a node to reach the required value to make sure the node sees all the updates it needs to process a transaction at question
3. Actually wait for a safeTime Delay Duration (DD) earlier than needed (to avoid waiting in most of the cases), compensate this by installing an update after DD in the future to make the timestamp math easier on the nodes

Using safe time

We must prevent a node processing a transaction at moment Now before the node has received all the schema updates with activation moment not later than Now. We are going to use Meta Storage for storing the schema updates (note that this implies that we use HLC from now on, not the physical clock that we used in the warm-up section). We are going to read schema updates from the Meta Storage on other nodes. To make sure we never read stale data (breaking AP), we could use the same approach that we use to make reads (in RO) from non-primary replicas: Safe Time.

Message M saying ‘Starting at moment T_u, schema is S’ is sent with timestamp T_m (that moves safeTime forward). A schema storage client that wants to obtain a schema active at moment T_op waits for Meta Storage’s safeTime>=T_op. T_u must be ≥ T_m (because we’ll construct messages in this way), so, if T_m<T_op, after waiting, the client sees M (as T_m<T_op<=safeTime, so T_m<safeTime, hence M is visible to the client’s node) and knows that S has been activated at T_u before it obtains schema for moment T_op.

This means that AP is maintained. Let’s suppose the opposite. This means that a client obtained a schema for some moment T_op>=T_u>=T_m while it was not knowing about M. But as the client had to wait for safeTime>=T_op+CS_max, T_{mmust have become ≤ than safeTime (even in the presence of clock skew), so the client must have seen message M. We have a contradiction.}

To maintain the ‘client sees his own schema changes’ property, we must make the client wait for CS_max after installing an update to make sure that Tu has become non-future on any node, so if the client issues a request to any node, they will see their update.

This approach guarantees correctness (AP is always maintained), but every node (that is not the Meta Storage leader), while trying to obtain a schema for the moment T_op=Now, will have to wait for safeTime visible on this node to reach T_op, which will take PD (propagation duration). So each transaction will be delayed for PD (on average). This creates a performance problem: latency increases.

Waiting for safe time in the past

Let’s make a couple of modifications.

1. Choose some delay duration. DD should be >= typical PD (so that we account for the network latency) and ≥ CS_max (so that clock skew does not make us wait).
2. When creating a message M telling the cluster about a schema update activation moment, choose the message timestamp T_m (moving safeTime forward) equal to Now, but assign T_u (activation moment) contained in that M to be T_m+DD
3. If a node wants to read the schema at T_op=Now, it will wait for safeTime≥T_op-DD instead of T_op and then it will take a schema at moment T_op. Most of the time, it will not need to wait, or the wait will be a lot shorter. (Please note that the subtraction only affects the safeTime we are waiting for, not the activation timestamp we look up after having waited; this makes it easier to reason about activation timestamps).

The downside is that schema change becomes visible only after some time. Even if a message is delivered faster than DD (or we are on the leader), we will still not see the schema until DD passes. We can alleviate this by returning the control to the user after additional time (DD) after the schema change is applied by the user, so the user that created the schema update will not be able to see the system in an inconsistent state (that is, not being able to see the change they have just introduced).

Here is a diagram that exemplifies the mechanism with the cases of 2 nodes: one healthy (that does not need to wait), another lagging (that has to wait).

Image Added

Risks and Assumptions

Schema updates are not visible right away

Even on the Meta -storage 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.

...

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 Storage leader. This means that the whole cluster is in an uncontrollable state. In such a situation, we become unable to process transactions.

...

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.

...

Reference Links

// N/A

Tickets