Back to Cockroach

Summary

docs/RFCS/20160425_drain_modes.md

26.1.312.0 KB
Original Source

Summary

Propose a draining process for a CockroachDB node to perform a graceful shutdown.

This draining process will be composed of two reduced modes of operation that will be run in parallel:

  • drain-clients mode: the server lets on-going SQL clients finish up to some deadline when context cancellation and cleanup is performed, and then politely refuses new work at the gateway in the manner which popular load-balancing SQL clients can handle best.
  • drain-leases mode: all range leases are transferred away from the node, new ones are not granted (in turn disabling most queues and active gossipping) and the draining node will not be a target for lease or range transfers. The draining node will decline any preemptive snapshots.

These modes are not related to the existing Stopper-based functionality and can be run independently (i.e. a server can temporarily run in drain-clients mode).

Motivation

In our Stopper usage, we've taken to fairly ruthlessly shutting down service of many components once (*Stopper).Quiesce() is called. In code, this manifests itself through copious use of the (*Stopper).ShouldQuiesce() channel even when not running inside of a task (or running long-running operations inside tasks).

This was motivated mainly by endless amounts of test failures around leaked goroutines or deadlocking operations and has served us reasonably well, keeping our moving parts in check.

However, now that we're looking at clusters and their maintenance, this simple "drop-all" approach isn't enough. See, for example #6198, #6197 or #5279. This RFC outlines proposed changes to accommodate use cases such as

  • clean shutdown: instead of dropping everything on the floor, politely block new clients and finish open work up to some deadline, transfer/release all leases to avoid a period of unavailability, and only then initiate shutdown via stopper.Stop(). This clean shutdown avoids:
    • Blocking schema changes on the order of minutes because table descriptor leases are not properly released #9493
    • Increased per-range activity caused by nodes trying to pick up expired epoch-based range leases.
    • Leaving around intents from ongoing transactions.
  • draining data off a node (i.e. have all or some subset of ranges migrate away, and wait until that has happened). Used for decommissioning, but could also be used to change hard-drives cleanly (i.e. drain data off a store, clean shutdown, switch hdd, start).
  • decommission (permanently remove), which is "most of the above" with the addition of announcing the downtime as forever to the cluster.
  • drain for update/migration (see the {F,Unf}reeze proposal in #6166): Essentially here we'll want to have the cluster in drain-clients mode and then propose a Freeze command on all Ranges.

This work will also result in

  • Auditing and improving our cancellation mechanism for individual queries.
  • Improving compatibility with server health-checking mechanisms used by third party proxies.

Detailed design

Implement the following:

go
// Put the SQL server into drain-clients mode within deadline, cancelling
// connections as necessary. A zero deadline undoes the effects of any
// prior drain operation.
(*pgwire.Server).DrainClients(deadline time.Time) error

// Put the Node into drain-lease mode if requested, or undo any previous
// mode change.
(*server.Node).DrainLeases(bool) error

SQL/drain-clients mode:

The v3conn struct will be extended to point to the server. Before and after reading messages from the client, a v3conn will check the draining status on the server. If both draining is set to true and v3conn.session.TxnState.State is NoTxn, the v3conn will exit the read loop, send an appropriate error code, and close the connection, thereby denying the reading and execution of statements that aren't part of an ongoing transaction. The v3conn will also be made aware of cancellation of its session's context which will be handled in the same way.

All active sessions will be found through the session registry (#10317). Sessions will be extended with a done channel which will be used by pgwire.Server to listen for the completion of sessions. Sessions will be created with a context derived from a cancellable parent context, thus offering a single point from which to cancel all sessions.

Additionally, pgwire.Server will not create new sessions once in draining mode.

After draining behavior is enabled, we have three classes of clients characterized by the behavior of the server:

  • Blocked in the read loop with no active transaction. No execution of any statements is ongoing.
  • Blocked in the read loop with an active transaction. No execution of any statements is ongoing but more are expected.
  • Not blocked in the read loop. The execution of (possibly a batch) of statements is ongoing.

A first pass through the registry will collect all channels that belong to sessions of the second and third class of clients. These clients will be given a deadline of drainMaxWait (default of 10s) to finish any work. Note that once work is complete, the v3conn will not block in the read loop and will instead exit. Once drainMaxWait has elapsed or there are no more sessions with active transactions, the parent context of all sessions will be canceled. Since a derived context is used to send RPCs, these RPCs will be canceled, resulting in the propagation of errors back to the Executor and the client. Additionally, plumbing will have to be added to important local work done on nodes so that this context cancellation leads to the interruption of this work.

pgwire.Server will listen for the completion of the remaining sessions up to a timeout of 1s. Some sessions might keep going indefinitely despite a canceled context.

The final stage is to delete table descriptor leases from the system.lease table to avoid future schema changes blocking. LeaseManager will be extended to delete all leases from system.lease with a refcount of 0. Since there might still be ongoing sessions, LeaseManager will enter a draining mode in which if any lease's refcount is decremented to 0, the lease will be deleted from system.lease. If sessions are still ongoing after this point, a log message would warn of this fact.

Note on closing SQL connections

The load balancing solutions for Postgres that seem to be the most popular are PGPool and HAProxy. Both systems have health check mechanisms that try to establish a SQL connection[1] [2]. HAProxy also supports doing only generic TCP checks. Because of how these health checks are performed, sending back a SQL error during the establishment of a connection will result in the draining node being correctly marked as down.

For failures during an existing or new session, PGPool has a fail_over_on_backend_error option which is triggered when

57P01 ADMIN SHUTDOWN admin_shutdown

is received from the backend. PGPool will retain session information under some conditions[3] and can reconnect transparently to a different backend.

HAProxy does not have similar functionality and both SQL/TCP errors are forwarded to the client.

We should therefore return

57P01 ADMIN SHUTDOWN admin_shutdown

to reject new connections and close existing ones for compatibility with PGPool's automatic failover and both PGPool's and HAProxy's health checks. However, clients should retry in the case that HAProxy is being used and want transparent failover when the establishment of a connection errors out.

Node/drain-leases mode:

(*server.Node).DrainLeases(bool) iterates over its store list and delegates to all contained stores. (*Replica).redirectOnOrAcquireLease checks with its store before requesting a new or extending an existing lease.

The Liveness proto will be extended with a draining field which will be taken into account in Allocator.TransferLeaseTarget and Allocator.AllocateTarget, resulting in no leases or ranges being transferred to a node that is known to be draining. Updating the node liveness record will trigger a gossip of the node's draining mode.

Transfers to a draining node could still happen before the gossiped Liveness has reached the source of the transfer. This will only be handled in the case of range transfers.

A draining node will decline a snapshot in HandleSnapshot if its store is draining. In the case of lease transfers, since they are proposed and applied as raft commands, there is no way for the recipient to reject a lease.

[NB: previously this RFC proposed modifying the lease transfer mechanism so that a draining node could immediately send back any leases transferred to it while draining. Since the draining process is a best effort to avoid unavailability and it's not sure this addition would be necessary or produce significant benefits, the modification of the lease transfer mechanism has been left out but could be implemented in the future if necessary]

To decrease the probability of unavailability, an optional 10s timeout will be introduced to wait for the gossip of the draining node's Liveness to propagate. This option will be off by default and operators will have the option of turning it on to prioritize availability over shutdown speed.

After this step, leases that a node's replicas currently hold will be transferred away using AdminTransferLease(target) where target will be found using Allocator.TransferLeaseTarget.

To allow commands that were sent to the node's replicas to complete, the draining node will wait for its replicas' command queues to drain up to a timeout of 1s. This timeout is necessary in the case that new leases are transferred to the node and it receives new commands.

Server/adminServer:

go

type DrainMode int
const (
  DrainClient DrainMode = 1 << iota
  DrainLeases
)

// For example, `s.Drain(DrainClient | DrainLeases)`
(*server.Server).Drain(mode DrainMode) error

and hook it up from (*adminServer).handleQuit (which currently has no way of accessing *Server, only *Node, so a shortcut may be taken for the time being if that seems opportune.

Drawbacks

Alternatives

  • Don't offer a drainMaxWait timeout to ongoing transactions. The idea of the timeout is to allow clients a grace period in which to complete work. The issue is that this timeout is arbitrary and it might make more sense to forcibly close these connections. It might also make sense to offer operators to set this timeout via an environment variable.
  • Reacquire table leases when the node goes back up. This was suggested in #9493 but does not fix the issue of having to wait for the lease to expire if the node does not come back up.

Future work

  • Move ranges to another node if we're draining the node for decommissioning or reject a shut down if not doing so would cause Raft groups to drop below quorum.
  • Change the lease transfer mechanism so a transferrer can transfer its timestamp cache's high water mark which would act as the low water mark of the recipient's timestamp cache. This is conditioned on not inserting reads in the command queue.

Unresolved questions