[tarantool-patches] [PATCH 1/1] rfc: describe box.ctl.promote protocol
Vladislav Shpilevoy
v.shpilevoy at tarantool.org
Wed Apr 18 23:52:48 MSK 2018
Part of #3055
---
Original: https://github.com/tarantool/tarantool/blob/gh-3055-box-ctl-promote-rfc/doc/rfc/3055-box_ctl_promote.md
Branch: https://github.com/tarantool/tarantool/tree/gh-3055-box-ctl-promote-rfc
Issue: https://github.com/tarantool/tarantool/issues/3055
doc/rfc/3055-box_ctl_promote.md | 280 ++++++++++++++++++++++++++++++++++
doc/rfc/3055-box_ctl_promote_img1.svg | 2 +
2 files changed, 282 insertions(+)
create mode 100644 doc/rfc/3055-box_ctl_promote.md
create mode 100644 doc/rfc/3055-box_ctl_promote_img1.svg
diff --git a/doc/rfc/3055-box_ctl_promote.md b/doc/rfc/3055-box_ctl_promote.md
new file mode 100644
index 000000000..9ce233203
--- /dev/null
+++ b/doc/rfc/3055-box_ctl_promote.md
@@ -0,0 +1,280 @@
+# Replicaset master promotion
+
+* **Status**: In progress
+* **Start date**: 02-03-2018
+* **Authors**: Vladislav Shpilevoy @Gerold103 \<v.shpilevoy at tarantool.org\>,
+Konstantin Osipov @kostja \<kostja at tarantool.org\>
+* **Issues**: [#3055](https://github.com/tarantool/tarantool/issues/3055),
+[#2625](https://github.com/tarantool/tarantool/issues/2625)
+
+## Summary
+
+Replicaset master promotion is a procedure of atomic making one slave be new
+master, and an old master be slave in a fullmesh master-slave replicaset. Master
+is a replica in read-write mode. Slave is a replica in read-only mode.
+
+Master promotion has API:
+```Lua
+--
+-- Called on a slave promotes its role to master, demoting an old
+-- one to slave. Called on a master returns an error.
+-- @param opts Options.
+-- * timeout - the time in which a promotion must be
+-- finished;
+-- * quorum - before an old master demotion its data must
+-- be synced with no less than quorum slave count,
+-- including the being promoted one;
+-- * force - in any case make the current slave be master
+-- even if an old one is unavailable, or quorum is not
+-- satisfied, or another promotion is detected.
+--
+-- @retval true Promotion is started.
+-- @retval nil, error Can not start promotion.
+--
+box.ctl.promote(opts)
+
+--
+-- Status of the latest finished or the currently working
+-- promotion round.
+-- @retval nil Promote() was not called since the instance has
+-- been started, or it was started on another instance,
+-- that could not sent promotion info to the current
+-- instance.
+-- @retval status A table with the format:
+-- {
+-- round_uuid = <Promotion round UUID, generated on
+-- initiator side>,
+-- promote_uuid = <UUID of the promotion initiator>,
+-- demote_uuid = <UUID of the old master>,
+-- state = <Human readable status of the algorithm - it
+-- can be finished ok, finished with an error,
+-- not finished being on one of algorithm steps>,
+-- step_number = <Promotion round step identifier>,
+-- error = <If the promotion is finished with an error,
+-- then here the error object is stored>,
+-- is_finished = <True, if the promotion round is
+-- finished>,
+-- start_ts = <Time of the promotion start on initiator
+-- clock>,
+-- update_ts = <Time of the last update of this promotion
+-- round by last sender clock>,
+-- end_ts = <Time of the promotion finish on initiator
+-- clock, if it is finished>,
+-- timeout = <Timeout of the promotion round>,
+-- quorum = <Requested quorum>,
+-- }
+--
+box.ctl.promotion_status()
+
+--
+-- Try to cancel the promotion.
+-- @retval true The promotion is canceled ok.
+-- @retval nil, error The promotion was not canceled.
+--
+box.ctl.promotion_cancel()
+
+--
+-- Remove info about all promotions from the entire cluster. It
+-- can be useful, when it is necessary to use a role specified in
+-- box.cfg{} even if it contradicts with a promotion result.
+--
+box.ctl.promotion_reset()
+```
+
+## Background and motivation
+
+The promote procedure strongly simplifies life of developers since they must not
+do all of the promotion steps manually, that in a common case is not a trivial
+task, as you can see in the algorithm description in the next section.
+
+The common algorithm, disregarding failures and their processing consists of the
+following steps:
+1. On an old master stop accepting DDL/DML - only DQL;
+2. Wait until all master data is received by needed slave count, including the
+new master candidate;
+3. Make the old master be slave;
+4. Make the slave be new master;
+5. Notify all other slaves, that master is changed.
+
+All of the steps are persisted in WAL, that guarantees, that even after a
+promotion participant is restarted, after waking up it will not forgot about
+promotion. Persistency eliminates any possibility of making the cluster have two
+masters after the promotion.
+
+## Detailed design
+
+Each cluster member has a special system space to distribute promotion steps
+over the cluster - `_promotion`:
+```Lua
+format = {}
+-- UUID of the promotion round, generated on an initiator.
+format[1] = {'round_uuid', 'string'}
+-- UUID of the sender instance.
+format[2] = {'source_uuid', 'string'}
+-- Increasing step identifier. It grows from 1 to the last one
+-- during promotion progress.
+format[3] = {'step_number', 'unsigned'}
+-- Timestamp, set by a sender using its own clock.
+format[4] = {'ts', 'unsigned'}
+--
+-- Type is what the sender want to get or send. Value depends on
+-- type.
+-- 'begin' - the first message that an initiator sends. Value
+-- contains all of the info, described in
+-- box.ctl.promotion_status() as a map.
+--
+-- 'status' - the message, sent by all of the cluster memebers on
+-- 'begin'. Value is either {role = 'master'} or
+-- {role = 'slave'}.
+--
+-- 'sync' - the message, that triggers an old master to enter
+-- read-only mode and sync with slaves. Value is nil.
+--
+-- 'success' - the next to last message, sent by an old master
+-- when sync was ok. After this the old master is
+-- demoted with no timeout. Value is nil.
+--
+-- 'error' - an error, that can be send by any cluster member.
+-- For example, it can be failed sync, or an existing
+-- promotion is found. Value is the error description.
+--
+-- 'commit' - the message sent by an initiator, when the
+-- 'success' status is replicated over quorum
+-- replicas. Value is nil.
+--
+format[5] = {'type', 'string'}
+format[6] = {'value', 'any', is_nullable = true}
+
+s = box.schema.create_space('_promotion', {format = format})
+```
+To participate in a promotion a cluster member just writes into `_promotion`
+space and waits until the record is replicated. This space is cleared by a
+garbage collector from finished promotions - it is ones with error or commited
+status. Only latest promotion is not deleted to be able to restore role after
+recovery.
+
+Below the protocol is described. On the image the state machine is showed:
+
+
+In the simplest case the being promoted instance is master already - immediately
+finish promotion with the error and with no persisting that. Now assume
+promote() is called on a slave. At first, the initiator broadcasts initial
+promotion status: `promote_uuid, step_number, start_ts, timeout, round_uuid,
+...`.
+
+Each cluster member, received the promotion status, checks if it already knows
+about another active promotions. If has, then responds error to the newer
+promotion request. Else it broadcasts its status that consists of role.
+
+Initiator collects responses from `_promotion` space. If an active promotion
+error is found, then stop the promotion - it is failed. Now assume active
+promotions are not found. The space `_promotion` looks like this:
+```YAML
+---
+- - ['<round_uuid>', '<init_uuid>', 0, <ts1>, 'begin', <status map>]
+ - ['<round_uuid>', '<slave1_uuid>', 0, <ts2>, 'status', {role = slave}]
+ - ['<round_uuid>', '<slave2_uuid>', 0, <ts3>, 'status', {role = slave}]
+ ...
+ - ['<round_uuid>', '<master_uuid>', 0, <ts4>, 'status', {role = master}]
+...
+```
+
+Consider role responses:
+* multiple masters are found;
+* a master is not found;
+* single master is found.
+
+### Multiple masters are found
+
+Immediately abort the promotion - promote must not destroy master-master
+replication. Broadcast this fact.
+
+### Master is not found
+
+It is possible in two cases: the cluster is read-only and actually does not have
+a master; a master is unavailable for the promotion initiator. These two cases
+can be controlled by quorum - if it is necessary to demote an old master and it
+is known to be active, the quorum must be equal to the cluster size. Consider
+the algorithm steps, when a master is not found.
+
+1. Initiator broadcasts sync request and waits timeout for sync with at least
+quorum replicas. On timeout broadcast `error` status.
+
+2. On success sync the initiator broadcasts the `success` status and enters
+master role. Else broadcast error status. When `success` is replicated over the
+quorum, the initiator broadcasts `commit`.
+
+The reason why the promotion must work with not found old master is that a user
+on failed promotion that left the cluster in read-only state must be able to
+return a master.
+
+### Single master is found
+
+It is possible in two cases as well: the cluster actually has a single master,
+or another master is unavailable. Necessity to ignore possibility of unavailable
+masters existence can be controlled by quorum. Consider the algorithm steps.
+
+1. Broadcast `sync` request. The old master interprets it as the demotion
+request. On demote the old master enters read-only mode on the at most current
+promotion round duration, and waits timeout for sync with quorum replicas
+including the requester. If the `sync` was ok, the `success` is broadcasted and
+the old master finishes its demotion becoming a slave. Else the `error` is
+broadcasted.
+
+2. The initiator receives sync result via waiting for new data in space
+`_promote`. If time is out, then just broadcast the `error` and finish. If sync
+result is received and it is failed, then the promotion is aborted already by
+this `error`. Assume the sync was ok. The initiator received `success` becomes a
+new master, entering the read-write mode. According to this schema, it is
+possible, that `error` can be written after `success`, if the old master
+broadcasted `success`, but it was replicated to the initiator too long. In such
+a case when this `error` reaches the old master, it must try to return to the
+master role. For this he checks if there no more promotions were runned, and
+just becomes a master back.
+
+3. After the `success` is broadcasted over quorum replicas, the initiator writes
+`commit` into `_promotion` space.
+
+### Promotion canceling
+
+Promotion cancel can be run from an old master and from an initiator. It just
+broadcasts `error` status via `_promotion` space, if an active promotion is
+found, and it is not already `success`, `error` or `commit`.
+
+### Recovery
+
+Recovery procedure consists of several independent cases, if a `_promotion`
+space is not empty:
+* Recovery of non-participant slave replica. Just do nothing.
+* Recovery of non-participant master replica. Ignore 'master' role - another
+master exists already.
+* Recovery of the old master.
+
+ Assume the found promotion state is `begin` or `status` or `sync` -
+ broadcast an `error` and become a master.
+
+ Assume the state is `error` - then the promotion is failed, and the
+ current replica is still a master.
+
+ Assume the `success` is found. Then the replica can not decide what to
+ do with no the promotion initiator. Indeed, it is possible, that the
+ initiator already became a new master, and broadcasted `commit`, that
+ just did not manage to reach the current replica. Even timeout can not
+ be used here, because the replica can be separated from a new master due
+ to network problems. So it becomes slave until a promotion result is got
+ from the initiator.
+
+ Assume the `commit` is found - then become a slave regardless of
+ configuration.
+
+* Recovery of the promotion initiator.
+
+ Assume the found promotion state is `begin` or `status` or `sync` -
+ broadcast an `error` and become a slave.
+
+ Assume the state is `error` - then become a slave regardless of
+ configuration.
+
+ Assume the status is `success` or `commit` - then become a master
+ regardless of configuration. If the status is `success` that is
+ replicated over quorum then broadcast `commit`.
More information about the Tarantool-patches
mailing list