returned by Ra, e.g. when a replica cannot be
restarted because of a concurrent delete
or because a QQ was inserted into a schema data
store but not yet registered as a process on
the node.
References #12013.
[Why]
Ra consistent queries are currently fragile in the sense that the query
function may run on a remote node and the function reference or MFA may
not be valid on that node. See previous commit for more details.
[How]
We perform local queries in `rabbit_db_maintenance:get_consistent/1`
when Khepri is enabled. This violates what the expectation from this
API, that's why it is a temporary measure, until a proper solution is
found.
[Why]
Mnesia is a very powerful and convenient tool for Erlang applications:
it is a persistent disc-based database, it handles replication accross
multiple Erlang nodes and it is available out-of-the-box from the
Erlang/OTP distribution. RabbitMQ relies on Mnesia to manage all its
metadata:
* virtual hosts' properties
* intenal users
* queue, exchange and binding declarations (not queues data)
* runtime parameters and policies
* ...
Unfortunately Mnesia makes it difficult to handle network partition and,
as a consequence, the merge conflicts between Erlang nodes once the
network partition is resolved. RabbitMQ provides several partition
handling strategies but they are not bullet-proof. Users still hit
situations where it is a pain to repair a cluster following a network
partition.
[How]
@kjnilsson created Ra [1], a Raft consensus library that RabbitMQ
already uses successfully to implement quorum queues and streams for
instance. Those queues do not suffer from network partitions.
We created Khepri [2], a new persistent and replicated database engine
based on Ra and we want to use it in place of Mnesia in RabbitMQ to
solve the problems with network partitions.
This patch integrates Khepri as an experimental feature. When enabled,
RabbitMQ will store all its metadata in Khepri instead of Mnesia.
This change comes with behavior changes. While Khepri remains disabled,
you should see no changes to the behavior of RabbitMQ. If there are
changes, it is a bug. After Khepri is enabled, there are significant
changes of behavior that you should be aware of.
Because it is based on the Raft consensus algorithm, when there is a
network partition, only the cluster members that are in the partition
with at least `(Number of nodes in the cluster ÷ 2) + 1` number of nodes
can "make progress". In other words, only those nodes may write to the
Khepri database and read from the database and expect a consistent
result.
For instance in a cluster of 5 RabbitMQ nodes:
* If there are two partitions, one with 3 nodes, one with 2 nodes, only
the group of 3 nodes will be able to write to the database.
* If there are three partitions, two with 2 nodes, one with 1 node, none
of the group can write to the database.
Because the Khepri database will be used for all kind of metadata, it
means that RabbitMQ nodes that can't write to the database will be
unable to perform some operations. A list of operations and what to
expect is documented in the associated pull request and the RabbitMQ
website.
This requirement from Raft also affects the startup of RabbitMQ nodes in
a cluster. Indeed, at least a quorum number of nodes must be started at
once to allow nodes to become ready.
To enable Khepri, you need to enable the `khepri_db` feature flag:
rabbitmqctl enable_feature_flag khepri_db
When the `khepri_db` feature flag is enabled, the migration code
performs the following two tasks:
1. It synchronizes the Khepri cluster membership from the Mnesia
cluster. It uses `mnesia_to_khepri:sync_cluster_membership/1` from
the `khepri_mnesia_migration` application [3].
2. It copies data from relevant Mnesia tables to Khepri, doing some
conversion if necessary on the way. Again, it uses
`mnesia_to_khepri:copy_tables/4` from `khepri_mnesia_migration` to do
it.
This can be performed on a running standalone RabbitMQ node or cluster.
Data will be migrated from Mnesia to Khepri without any service
interruption. Note that during the migration, the performance may
decrease and the memory footprint may go up.
Because this feature flag is considered experimental, it is not enabled
by default even on a brand new RabbitMQ deployment.
More about the implementation details below:
In the past months, all accesses to Mnesia were isolated in a collection
of `rabbit_db*` modules. This is where the integration of Khepri mostly
takes place: we use a function called `rabbit_khepri:handle_fallback/1`
which selects the database and perform the query or the transaction.
Here is an example from `rabbit_db_vhost`:
* Up until RabbitMQ 3.12.x:
get(VHostName) when is_binary(VHostName) ->
get_in_mnesia(VHostName).
* Starting with RabbitMQ 3.13.0:
get(VHostName) when is_binary(VHostName) ->
rabbit_khepri:handle_fallback(
#{mnesia => fun() -> get_in_mnesia(VHostName) end,
khepri => fun() -> get_in_khepri(VHostName) end}).
This `rabbit_khepri:handle_fallback/1` function relies on two things:
1. the fact that the `khepri_db` feature flag is enabled, in which case
it always executes the Khepri-based variant.
4. the ability or not to read and write to Mnesia tables otherwise.
Before the feature flag is enabled, or during the migration, the
function will try to execute the Mnesia-based variant. If it succeeds,
then it returns the result. If it fails because one or more Mnesia
tables can't be used, it restarts from scratch: it means the feature
flag is being enabled and depending on the outcome, either the
Mnesia-based variant will succeed (the feature flag couldn't be enabled)
or the feature flag will be marked as enabled and it will call the
Khepri-based variant. The meat of this function really lives in the
`khepri_mnesia_migration` application [3] and
`rabbit_khepri:handle_fallback/1` is a wrapper on top of it that knows
about the feature flag.
However, some calls to the database do not depend on the existence of
Mnesia tables, such as functions where we need to learn about the
members of a cluster. For those, we can't rely on exceptions from
Mnesia. Therefore, we just look at the state of the feature flag to
determine which database to use. There are two situations though:
* Sometimes, we need the feature flag state query to block because the
function interested in it can't return a valid answer during the
migration. Here is an example:
case rabbit_khepri:is_enabled(RemoteNode) of
true -> can_join_using_khepri(RemoteNode);
false -> can_join_using_mnesia(RemoteNode)
end
* Sometimes, we need the feature flag state query to NOT block (for
instance because it would cause a deadlock). Here is an example:
case rabbit_khepri:get_feature_state() of
enabled -> members_using_khepri();
_ -> members_using_mnesia()
end
Direct accesses to Mnesia still exists. They are limited to code that is
specific to Mnesia such as classic queue mirroring or network partitions
handling strategies.
Now, to discover the Mnesia tables to migrate and how to migrate them,
we use an Erlang module attribute called
`rabbit_mnesia_tables_to_khepri_db` which indicates a list of Mnesia
tables and an associated converter module. Here is an example in the
`rabbitmq_recent_history_exchange` plugin:
-rabbit_mnesia_tables_to_khepri_db(
[{?RH_TABLE, rabbit_db_rh_exchange_m2k_converter}]).
The converter module — `rabbit_db_rh_exchange_m2k_converter` in this
example — is is fact a "sub" converter module called but
`rabbit_db_m2k_converter`. See the documentation of a `mnesia_to_khepri`
converter module to learn more about these modules.
[1] https://github.com/rabbitmq/ra
[2] https://github.com/rabbitmq/khepri
[3] https://github.com/rabbitmq/khepri_mnesia_migration
See #7206.
Co-authored-by: Jean-Sébastien Pédron <jean-sebastien@rabbitmq.com>
Co-authored-by: Diana Parra Corbacho <dparracorbac@vmware.com>
Co-authored-by: Michael Davis <mcarsondavis@gmail.com>
So far, we had the following functions to list nodes in a RabbitMQ
cluster:
* `rabbit_mnesia:cluster_nodes/1` to get members of the Mnesia cluster;
the argument was used to select members (all members or only those
running Mnesia and participating in the cluster)
* `rabbit_nodes:all/0` to get all members of the Mnesia cluster
* `rabbit_nodes:all_running/0` to get all members who currently run
Mnesia
Basically:
* `rabbit_nodes:all/0` calls `rabbit_mnesia:cluster_nodes(all)`
* `rabbit_nodes:all_running/0` calls `rabbit_mnesia:cluster_nodes(running)`
We also have:
* `rabbit_node_monitor:alive_nodes/1` which filters the given list of
nodes to only select those currently running Mnesia
* `rabbit_node_monitor:alive_rabbit_nodes/1` which filters the given
list of nodes to only select those currently running RabbitMQ
Most of the code uses `rabbit_mnesia:cluster_nodes/1` or the
`rabbit_nodes:all*/0` functions. `rabbit_mnesia:cluster_nodes(running)`
or `rabbit_nodes:all_running/0` is often used as a close approximation
of "all cluster members running RabbitMQ". This list might be incorrect
in times where a node is joining the clustered or is being worked on
(i.e. Mnesia is running but not RabbitMQ).
With Khepri, there won't be the same possible approximation because we
will try to keep Khepri/Ra running even if RabbitMQ is stopped to
expand/shrink the cluster.
So in order to clarify what we want when we query a list of nodes, this
patch introduces the following functions:
* `rabbit_nodes:list_members/0` to get all cluster members, regardless
of their state
* `rabbit_nodes:list_reachable/0` to get all cluster members we can
reach using Erlang distribution, regardless of the state of RabbitMQ
* `rabbit_nodes:list_running/0` to get all cluster members who run
RabbitMQ, regardless of the maintenance state
* `rabbit_nodes:list_serving/0` to get all cluster members who run
RabbitMQ and are accepting clients
In addition to the list functions, there are the corresponding
`rabbit_nodes:is_*(Node)` checks and `rabbit_nodes:filter_*(Nodes)`
filtering functions.
The code is modified to use these new functions. One possible
significant change is that the new list functions will perform RPC calls
to query the nodes' state, unlike `rabbit_mnesia:cluster_nodes(running)`.
When a node gets drained (i.e. goes into maintenance mode), only local
connections should be terminated.
However, prior to this commit, all MQTT connections got terminated
cluster-wide when a single node was drained.
Maintenance mode, introduced in RabbitMQ 3.8.x, was a breaking change
protected behind a feature flag. This allowed a RabbitMQ cluster to be
upgraded one node at a time, without having to stop the entire cluster.
The compatibility code is in the wild for long enough. The
`maintenance_mode_status` feature flag was marked as required in a
previous commit (see #5202). This allows us to remove code in this
patch.
References #5215.
The code of `migrate_leadership_to_existing_replica` previously assumed
that it can control the target node of a failover of mirrored queues.
While its sibling function `transfer_leadership` can do that (as it drops all
mirrors besides the one on the target node)
`migrate_leadership_to_existing_replica` tries to be less intrusive and
only tries to migrate the queue away for the current primary.
However it then used `wait_for_new_master` to wait for the queue to actually
transfer the primary to the specified target node. As the specified target
node might not ever become the primary of the queue (as this decission is
only made between the remaining mirrors) this can cause the migrate operation
to run into a timeout (10 sec per default).
As `migrate_leadership_to_existing_replica` is only used during
`transfer_leadership_of_classic_mirrored_queues` its only goal is to get the
primary away from the current node. Therefor we can just wait for the
queue to become active on some other node instead of expecting a specific
node to become the primary.
Adds WORKSPACE.bazel, BUILD.bazel & *.bzl files for partial build & test with Bazel. Introduces a build-time dependency on https://github.com/rabbitmq/bazel-erlang
The time this operation can take in clusters with a lot of classic
mirrored queue (say, 10s or 100s of thousands) be prohibitive for
upgrades.
Upgrades that use a health check to ensure that there are in-sync
replicas before entering maintenance mode, in which case
the transfer is not really necessary.
All of the above is more obvious with the recent changes in #2749.
Since we only consider nodes hosting
in-sync replicas for transfer candidates,
we can drop only one mirror instead of N,
and reduce the load caused by this operation.
This does not affect CMQ leadership
transfer when performed in the context
'rabbitmq-queues rebalance'
Pair: @dcorbacho, @mkuratczyk
of classic mirrored queues.
There are cases when asynchronously adding mirrors makes
a lot of sense: e.g. when a new node joins the cluster.
In this case, if we add mirrors asynchronously, this
operation will race with the step that removes mirrors.
As a result, we can end up with a queue that decided
that it had no promotable replicas => data loss
from the transfer.
Closes#2749.
Pairs: @dcorbacho, @mkuratczyk
Michal Kuratczyk
Iliia Khaprov
Michael Klishin
Karl Nilsson
Michael Klishin
Diana Parra Corbacho
deterclosed
Jean-Sébastien Pédron
Michael Klishin
Alexey Lebedeff
Michael Klishin
Michal Kuratczyk
David Ansari
Luke Bakken
Jean-Sébastien Pédron
Felix Huettner
Philip Kuryloski
kjnilsson
Michal Kuratczyk