This check is expected to succeed and the status is expected to be
printed to stdout rather than stderr. This change silences the status
output. The status text was printed mistakenly previously because we
captured stderr rather than stdout.
This previously emitted a warning because Elixir will rebind `this_node`
by default, so the `this_node` binding in the line above was unused.
(As opposed to Erlang which would treat this as a match - rejecting
the binding if `this_node` was not equal to the value being matched.)
The node needed to be adjusted as well - `node()` returned the ExUnit
runner's node while the command returned the remote node, which is
stored in the context under `opts.node`.
--experimental is no longer particularly fair to Khepri,
which is not enabled by default because of its enormous
scope, and because once enabled, it cannot be disabled.
--opt-in would be a better name but --experimental
remains for backwards compatiblity.
When both are specified, we consider that the
user opts in if at least one of the flags is
set to true.
The spec of `rabbit_exchange:declare/7` needs to be updated to return
`{ok, Exchange} | {error, Reason}` instead of the old return value of
`rabbit_types:exchange()`. This is safe to do since `declare/7` is not
called by RPC - from the CLI or otherwise - outside of test suites, and
in test suites only through the CLI's `TestHelper.declare_exchange/7`.
Callers of this helper are updated in this commit.
Otherwise this commit updates callers to unwrap the `{ok, Exchange}`
and bubble up errors.
## What
Similar to Native MQTT in #5895, this commits implements Native AMQP 1.0.
By "native", we mean do not proxy via AMQP 0.9.1 anymore.
## Why
Native AMQP 1.0 comes with the following major benefits:
1. Similar to Native MQTT, this commit provides better throughput, latency,
scalability, and resource usage for AMQP 1.0.
See https://blog.rabbitmq.com/posts/2023/03/native-mqtt for native MQTT improvements.
See further below for some benchmarks.
2. Since AMQP 1.0 is not limited anymore by the AMQP 0.9.1 protocol,
this commit allows implementing more AMQP 1.0 features in the future.
Some features are already implemented in this commit (see next section).
3. Simpler, better understandable, and more maintainable code.
Native AMQP 1.0 as implemented in this commit has the
following major benefits compared to AMQP 0.9.1:
4. Memory and disk alarms will only stop accepting incoming TRANSFER frames.
New connections can still be created to consume from RabbitMQ to empty queues.
5. Due to 4. no need anymore for separate connections for publishers and
consumers as we currently recommended for AMQP 0.9.1. which potentially
halves the number of physical TCP connections.
6. When a single connection sends to multiple target queues, a single
slow target queue won't block the entire connection.
Publisher can still send data quickly to all other target queues.
7. A publisher can request whether it wants publisher confirmation on a per-message basis.
In AMQP 0.9.1 publisher confirms are configured per channel only.
8. Consumers can change their "prefetch count" dynamically which isn't
possible in our AMQP 0.9.1 implementation. See #10174
9. AMQP 1.0 is an extensible protocol
This commit also fixes dozens of bugs present in the AMQP 1.0 plugin in
RabbitMQ 3.x - most of which cannot be backported due to the complexity
and limitations of the old 3.x implementation.
This commit contains breaking changes and is therefore targeted for RabbitMQ 4.0.
## Implementation details
1. Breaking change: With Native AMQP, the behaviour of
```
Convert AMQP 0.9.1 message headers to application properties for an AMQP 1.0 consumer
amqp1_0.convert_amqp091_headers_to_app_props = false | true (default false)
Convert AMQP 1.0 Application Properties to AMQP 0.9.1 headers
amqp1_0.convert_app_props_to_amqp091_headers = false | true (default false)
```
will break because we always convert according to the message container conversions.
For example, AMQP 0.9.1 x-headers will go into message-annotations instead of application properties.
Also, `false` won’t be respected since we always convert the headers with message containers.
2. Remove rabbit_queue_collector
rabbit_queue_collector is responsible for synchronously deleting
exclusive queues. Since the AMQP 1.0 plugin never creates exclusive
queues, rabbit_queue_collector doesn't need to be started in the first
place. This will save 1 Erlang process per AMQP 1.0 connection.
3. 7 processes per connection + 1 process per session in this commit instead of
7 processes per connection + 15 processes per session in 3.x
Supervision hierarchy got re-designed.
4. Use 1 writer process per AMQP 1.0 connection
AMQP 0.9.1 uses a separate rabbit_writer Erlang process per AMQP 0.9.1 channel.
Prior to this commit, AMQP 1.0 used a separate rabbit_amqp1_0_writer process per AMQP 1.0 session.
Advantage of single writer proc per session (prior to this commit):
* High parallelism for serialising packets if multiple sessions within
a connection write heavily at the same time.
This commit uses a single writer process per AMQP 1.0 connection that is
shared across all AMQP 1.0 sessions.
Advantages of single writer proc per connection (this commit):
* Lower memory usage with hundreds of thousands of AMQP 1.0 sessions
* Less TCP and IP header overhead given that the single writer process
can accumulate across all sessions bytes before flushing the socket.
In other words, this commit decides that a reader / writer process pair
per AMQP 1.0 connection is good enough for bi-directional TRANSFER flows.
Having a writer per session is too heavy.
We still ensure high throughput by having separate reader, writer, and
session processes.
5. Transform rabbit_amqp1_0_writer into gen_server
Why:
Prior to this commit, when clicking on the AMQP 1.0 writer process in
observer, the process crashed.
Instead of handling all these debug messages of the sys module, it's better
to implement a gen_server.
There is no advantage of using a special OTP process over gen_server
for the AMQP 1.0 writer.
gen_server also provides cleaner format status output.
How:
Message callbacks return a timeout of 0.
After all messages in the inbox are processed, the timeout message is
handled by flushing any pending bytes.
6. Remove stats timer from writer
AMQP 1.0 connections haven't emitted any stats previously.
7. When there are contiguous queue confirmations in the session process
mailbox, batch them. When the confirmations are sent to the publisher, a
single DISPOSITION frame is sent for contiguously confirmed delivery
IDs.
This approach should be good enough. However it's sub optimal in
scenarios where contiguous delivery IDs that need confirmations are rare,
for example:
* There are multiple links in the session with different sender
settlement modes and sender publishes across these links interleaved.
* sender settlement mode is mixed and sender publishes interleaved settled
and unsettled TRANSFERs.
8. Introduce credit API v2
Why:
The AMQP 0.9.1 credit extension which is to be removed in 4.0 was poorly
designed since basic.credit is a synchronous call into the queue process
blocking the entire AMQP 1.0 session process.
How:
Change the interactions between queue clients and queue server
implementations:
* Clients only request a credit reply if the FLOW's `echo` field is set
* Include all link flow control state held by the queue process into a
new credit_reply queue event:
* `available` after the queue sends any deliveries
* `link-credit` after the queue sends any deliveries
* `drain` which allows us to combine the old queue events
send_credit_reply and send_drained into a single new queue event
credit_reply.
* Include the consumer tag into the credit_reply queue event such that
the AMQP 1.0 session process can process any credit replies
asynchronously.
Link flow control state `delivery-count` also moves to the queue processes.
The new interactions are hidden behind feature flag credit_api_v2 to
allow for rolling upgrades from 3.13 to 4.0.
9. Use serial number arithmetic in quorum queues and session process.
10. Completely bypass the rabbit_limiter module for AMQP 1.0
flow control. The goal is to eventually remove the rabbit_limiter module
in 4.0 since AMQP 0.9.1 global QoS will be unsupported in 4.0. This
commit lifts the AMQP 1.0 link flow control logic out of rabbit_limiter
into rabbit_queue_consumers.
11. Fix credit bug for streams:
AMQP 1.0 settlements shouldn't top up link credit,
only FLOW frames should top up link credit.
12. Allow sender settle mode unsettled for streams
since AMQP 1.0 acknowledgements to streams are no-ops (currently).
13. Fix AMQP 1.0 client bugs
Auto renewing credits should not be related to settling TRANSFERs.
Remove field link_credit_unsettled as it was wrong and confusing.
Prior to this commit auto renewal did not work when the sender uses
sender settlement mode settled.
14. Fix AMQP 1.0 client bugs
The wrong outdated Link was passed to function auto_flow/2
15. Use osiris chunk iterator
Only hold messages of uncompressed sub batches in memory if consumer
doesn't have sufficient credits.
Compressed sub batches are skipped for non Stream protocol consumers.
16. Fix incoming link flow control
Always use confirms between AMQP 1.0 queue clients and queue servers.
As already done internally by rabbit_fifo_client and
rabbit_stream_queue, use confirms for classic queues as well.
17. Include link handle into correlation when publishing messages to target queues
such that session process can correlate confirms from target queues to
incoming links.
18. Only grant more credits to publishers if publisher hasn't sufficient credits
anymore and there are not too many unconfirmed messages on the link.
19. Completely ignore `block` and `unblock` queue actions and RabbitMQ credit flow
between classic queue process and session process.
20. Link flow control is independent between links.
A client can refer to a queue or to an exchange with multiple
dynamically added target queues. Multiple incoming links can also fan
in to the same queue. However the link topology looks like, this
commit ensures that each link is only granted more credits if that link
isn't overloaded.
21. A connection or a session can send to many different queues.
In AMQP 0.9.1, a single slow queue will lead to the entire channel, and
then entire connection being blocked.
This commit makes sure that a single slow queue from one link won't slow
down sending on other links.
For example, having link A sending to a local classic queue and
link B sending to 5 replica quorum queue, link B will naturally
grant credits slower than link A. So, despite the quorum queue being
slower in confirming messages, the same AMQP 1.0 connection and session
can still pump data very fast into the classic queue.
22. If cluster wide memory or disk alarm occurs.
Each session sends a FLOW with incoming-window to 0 to sending client.
If sending clients don’t obey, force disconnect the client.
If cluster wide memory alarm clears:
Each session resumes with a FLOW defaulting to initial incoming-window.
23. All operations apart of publishing TRANSFERS to RabbitMQ can continue during cluster wide alarms,
specifically, attaching consumers and consuming, i.e. emptying queues.
There is no need for separate AMQP 1.0 connections for publishers and consumers as recommended in our AMQP 0.9.1 implementation.
24. Flow control summary:
* If queue becomes bottleneck, that’s solved by slowing down individual sending links (AMQP 1.0 link flow control).
* If session becomes bottleneck (more unlikely), that’s solved by AMQP 1.0 session flow control.
* If connection becomes bottleneck, it naturally won’t read fast enough from the socket causing TCP backpressure being applied.
Nowhere will RabbitMQ internal credit based flow control (i.e. module credit_flow) be used on the incoming AMQP 1.0 message path.
25. Register AMQP sessions
Prefer local-only pg over our custom pg_local implementation as
pg is a better process group implementation than pg_local.
pg_local was identified as bottleneck in tests where many MQTT clients were disconnected at once.
26. Start a local-only pg when Rabbit boots:
> A scope can be kept local-only by using a scope name that is unique cluster-wide, e.g. the node name:
> pg:start_link(node()).
Register AMQP 1.0 connections and sessions with pg.
In future we should remove pg_local and instead use the new local-only
pg for all registered processes such as AMQP 0.9.1 connections and channels.
27. Requeue messages if link detached
Although the spec allows to settle delivery IDs on detached links, RabbitMQ does not respect the 'closed'
field of the DETACH frame and therefore handles every DETACH frame as closed. Since the link is closed,
we expect every outstanding delivery to be requeued.
In addition to consumer cancellation, detaching a link therefore causes in flight deliveries to be requeued.
Note that this behaviour is different from merely consumer cancellation in AMQP 0.9.1:
"After a consumer is cancelled there will be no future deliveries dispatched to it. Note that there can
still be "in flight" deliveries dispatched previously. Cancelling a consumer will neither discard nor requeue them."
[https://www.rabbitmq.com/consumers.html#unsubscribing]
An AMQP receiver can first drain, and then detach to prevent "in flight" deliveries
28. Init AMQP session with BEGIN frame
Similar to how there can't be an MQTT processor without a CONNECT
frame, there can't be an AMQP session without a BEGIN frame.
This allows having strict dialyzer types for session flow control
fields (i.e. not allowing 'undefined').
29. Move serial_number to AMQP 1.0 common lib
such that it can be used by both AMQP 1.0 server and client
30. Fix AMQP client to do serial number arithmetic.
31. AMQP client: Differentiate between delivery-id and transfer-id for better
understandability.
32. Fix link flow control in classic queues
This commit fixes
```
java -jar target/perf-test.jar -ad false -f persistent -u cq -c 3000 -C 1000000 -y 0
```
followed by
```
./omq -x 0 amqp -T /queue/cq -D 1000000 --amqp-consumer-credits 2
```
Prior to this commit, (and on RabbitMQ 3.x) the consuming would halt after around
8 - 10,000 messages.
The bug was that in flight messages from classic queue process to
session process were not taken into account when topping up credit to
the classic queue process.
Fixes#2597
The solution to this bug (and a much cleaner design anyway independent of
this bug) is that queues should hold all link flow control state including
the delivery-count.
Hence, when credit API v2 is used the delivery-count will be held by the
classic queue process, quorum queue process, and stream queue client
instead of managing the delivery-count in the session.
33. The double level crediting between (a) session process and
rabbit_fifo_client, and (b) rabbit_fifo_client and rabbit_fifo was
removed. Therefore, instead of managing 3 separate delivery-counts (i. session,
ii. rabbit_fifo_client, iii. rabbit_fifo), only 1 delivery-count is used
in rabbit_fifo. This is a big simplification.
34. This commit fixes quorum queues without bumping the machine version
nor introducing new rabbit_fifo commands.
Whether credit API v2 is used is solely determined at link attachment time
depending on whether feature flag credit_api_v2 is enabled.
Even when that feature flag will be enabled later on, this link will
keep using credit API v1 until detached (or the node is shut down).
Eventually, after feature flag credit_api_v2 has been enabled and a
subsequent rolling upgrade, all links will use credit API v2.
This approach is safe and simple.
The 2 alternatives to move delivery-count from the session process to the
queue processes would have been:
i. Explicit feature flag credit_api_v2 migration function
* Can use a gen_server:call and only finish migration once all delivery-counts were migrated.
Cons:
* Extra new message format just for migration is required.
* Risky as migration will fail if a target queue doesn’t reply.
ii. Session always includes DeliveryCountSnd when crediting to the queue:
Cons:
* 2 delivery counts will be hold simultaneously in session proc and queue proc;
could be solved by deleting the session proc’s delivery-count for credit-reply
* What happens if the receiver doesn’t provide credit for a very long time? Is that a problem?
35. Support stream filtering in AMQP 1.0 (by @acogoluegnes)
Use the x-stream-filter-value message annotation
to carry the filter value in a published message.
Use the rabbitmq:stream-filter and rabbitmq:stream-match-unfiltered
filters when creating a receiver that wants to filter
out messages from a stream.
36. Remove credit extension from AMQP 0.9.1 client
37. Support maintenance mode closing AMQP 1.0 connections.
38. Remove AMQP 0.9.1 client dependency from AMQP 1.0 implementation.
39. Move AMQP 1.0 plugin to the core. AMQP 1.0 is enabled by default.
The old rabbitmq_amqp1_0 plugin will be kept as a no-op plugin to prevent deployment
tools from failing that execute:
```
rabbitmq-plugins enable rabbitmq_amqp1_0
rabbitmq-plugins disable rabbitmq_amqp1_0
```
40. Breaking change: Remove CLI command `rabbitmqctl list_amqp10_connections`.
Instead, list both AMQP 0.9.1 and AMQP 1.0 connections in `list_connections`:
```
rabbitmqctl list_connections protocol
Listing connections ...
protocol
{1, 0}
{0,9,1}
```
## Benchmarks
### Throughput & Latency
Setup:
* Single node Ubuntu 22.04
* Erlang 26.1.1
Start RabbitMQ:
```
make run-broker PLUGINS="rabbitmq_management rabbitmq_amqp1_0" FULL=1 RABBITMQ_SERVER_ADDITIONAL_ERL_ARGS="+S 3"
```
Predeclare durable classic queue cq1, durable quorum queue qq1, durable stream queue sq1.
Start client:
https://github.com/ssorj/quiverhttps://hub.docker.com/r/ssorj/quiver/tags (digest 453a2aceda64)
```
docker run -it --rm --add-host host.docker.internal:host-gateway ssorj/quiver:latest
bash-5.1# quiver --version
quiver 0.4.0-SNAPSHOT
```
1. Classic queue
```
quiver //host.docker.internal//amq/queue/cq1 --durable --count 1m --duration 10m --body-size 12 --credit 1000
```
This commit:
```
Count ............................................. 1,000,000 messages
Duration ............................................... 73.8 seconds
Sender rate .......................................... 13,548 messages/s
Receiver rate ........................................ 13,547 messages/s
End-to-end rate ...................................... 13,547 messages/s
Latencies by percentile:
0% ........ 0 ms 90.00% ........ 9 ms
25% ........ 2 ms 99.00% ....... 14 ms
50% ........ 4 ms 99.90% ....... 17 ms
100% ....... 26 ms 99.99% ....... 24 ms
```
RabbitMQ 3.x (main branch as of 30 January 2024):
```
---------------------- Sender ----------------------- --------------------- Receiver ---------------------- --------
Time [s] Count [m] Rate [m/s] CPU [%] RSS [M] Time [s] Count [m] Rate [m/s] CPU [%] RSS [M] Lat [ms]
----------------------------------------------------- ----------------------------------------------------- --------
2.1 130,814 65,342 6 73.6 2.1 3,217 1,607 0 8.0 511
4.1 163,580 16,367 2 74.1 4.1 3,217 0 0 8.0 0
6.1 229,114 32,767 3 74.1 6.1 3,217 0 0 8.0 0
8.1 261,880 16,367 2 74.1 8.1 67,874 32,296 8 8.2 7,662
10.1 294,646 16,367 2 74.1 10.1 67,874 0 0 8.2 0
12.1 360,180 32,734 3 74.1 12.1 67,874 0 0 8.2 0
14.1 392,946 16,367 3 74.1 14.1 68,604 365 0 8.2 12,147
16.1 458,480 32,734 3 74.1 16.1 68,604 0 0 8.2 0
18.1 491,246 16,367 2 74.1 18.1 68,604 0 0 8.2 0
20.1 556,780 32,767 4 74.1 20.1 68,604 0 0 8.2 0
22.1 589,546 16,375 2 74.1 22.1 68,604 0 0 8.2 0
receiver timed out
24.1 622,312 16,367 2 74.1 24.1 68,604 0 0 8.2 0
quiver: error: PlanoProcessError: Command 'quiver-arrow receive //host.docker.internal//amq/queue/cq1 --impl qpid-proton-c --duration 10m --count 1m --rate 0 --body-size 12 --credit 1000 --transaction-size 0 --timeout 10 --durable --output /tmp/quiver-otujr23y' returned non-zero exit status 1.
Traceback (most recent call last):
File "/usr/local/lib/quiver/python/quiver/pair.py", line 144, in run
_plano.wait(receiver, check=True)
File "/usr/local/lib/quiver/python/plano/main.py", line 1243, in wait
raise PlanoProcessError(proc)
plano.main.PlanoProcessError: Command 'quiver-arrow receive //host.docker.internal//amq/queue/cq1 --impl qpid-proton-c --duration 10m --count 1m --rate 0 --body-size 12 --credit 1000 --transaction-size 0 --timeout 10 --durable --output /tmp/quiver-otujr23y' returned non-zero exit status 1.
```
2. Quorum queue:
```
quiver //host.docker.internal//amq/queue/qq1 --durable --count 1m --duration 10m --body-size 12 --credit 1000
```
This commit:
```
Count ............................................. 1,000,000 messages
Duration .............................................. 101.4 seconds
Sender rate ........................................... 9,867 messages/s
Receiver rate ......................................... 9,868 messages/s
End-to-end rate ....................................... 9,865 messages/s
Latencies by percentile:
0% ....... 11 ms 90.00% ....... 23 ms
25% ....... 15 ms 99.00% ....... 28 ms
50% ....... 18 ms 99.90% ....... 33 ms
100% ....... 49 ms 99.99% ....... 47 ms
```
RabbitMQ 3.x:
```
---------------------- Sender ----------------------- --------------------- Receiver ---------------------- --------
Time [s] Count [m] Rate [m/s] CPU [%] RSS [M] Time [s] Count [m] Rate [m/s] CPU [%] RSS [M] Lat [ms]
----------------------------------------------------- ----------------------------------------------------- --------
2.1 130,814 65,342 9 69.9 2.1 18,430 9,206 5 7.6 1,221
4.1 163,580 16,375 5 70.2 4.1 18,867 218 0 7.6 2,168
6.1 229,114 32,767 6 70.2 6.1 18,867 0 0 7.6 0
8.1 294,648 32,734 7 70.2 8.1 18,867 0 0 7.6 0
10.1 360,182 32,734 6 70.2 10.1 18,867 0 0 7.6 0
12.1 425,716 32,767 6 70.2 12.1 18,867 0 0 7.6 0
receiver timed out
14.1 458,482 16,367 5 70.2 14.1 18,867 0 0 7.6 0
quiver: error: PlanoProcessError: Command 'quiver-arrow receive //host.docker.internal//amq/queue/qq1 --impl qpid-proton-c --duration 10m --count 1m --rate 0 --body-size 12 --credit 1000 --transaction-size 0 --timeout 10 --durable --output /tmp/quiver-b1gcup43' returned non-zero exit status 1.
Traceback (most recent call last):
File "/usr/local/lib/quiver/python/quiver/pair.py", line 144, in run
_plano.wait(receiver, check=True)
File "/usr/local/lib/quiver/python/plano/main.py", line 1243, in wait
raise PlanoProcessError(proc)
plano.main.PlanoProcessError: Command 'quiver-arrow receive //host.docker.internal//amq/queue/qq1 --impl qpid-proton-c --duration 10m --count 1m --rate 0 --body-size 12 --credit 1000 --transaction-size 0 --timeout 10 --durable --output /tmp/quiver-b1gcup43' returned non-zero exit status 1.
```
3. Stream:
```
quiver-arrow send //host.docker.internal//amq/queue/sq1 --durable --count 1m -d 10m --summary --verbose
```
This commit:
```
Count ............................................. 1,000,000 messages
Duration ................................................ 8.7 seconds
Message rate ........................................ 115,154 messages/s
```
RabbitMQ 3.x:
```
Count ............................................. 1,000,000 messages
Duration ............................................... 21.2 seconds
Message rate ......................................... 47,232 messages/s
```
### Memory usage
Start RabbitMQ:
```
ERL_MAX_PORTS=3000000 RABBITMQ_SERVER_ADDITIONAL_ERL_ARGS="+P 3000000 +S 6" make run-broker PLUGINS="rabbitmq_amqp1_0" FULL=1 RABBITMQ_CONFIG_FILE="rabbitmq.conf"
```
```
/bin/cat rabbitmq.conf
tcp_listen_options.sndbuf = 2048
tcp_listen_options.recbuf = 2048
vm_memory_high_watermark.relative = 0.95
vm_memory_high_watermark_paging_ratio = 0.95
loopback_users = none
```
Create 50k connections with 2 sessions per connection, i.e. 100k session in total:
```go
package main
import (
"context"
"log"
"time"
"github.com/Azure/go-amqp"
)
func main() {
for i := 0; i < 50000; i++ {
conn, err := amqp.Dial(context.TODO(), "amqp://nuc", &amqp.ConnOptions{SASLType: amqp.SASLTypeAnonymous()})
if err != nil {
log.Fatal("dialing AMQP server:", err)
}
_, err = conn.NewSession(context.TODO(), nil)
if err != nil {
log.Fatal("creating AMQP session:", err)
}
_, err = conn.NewSession(context.TODO(), nil)
if err != nil {
log.Fatal("creating AMQP session:", err)
}
}
log.Println("opened all connections")
time.Sleep(5 * time.Hour)
}
```
This commit:
```
erlang:memory().
[{total,4586376480},
{processes,4025898504},
{processes_used,4025871040},
{system,560477976},
{atom,1048841},
{atom_used,1042841},
{binary,233228608},
{code,21449982},
{ets,108560464}]
erlang:system_info(process_count).
450289
```
7 procs per connection + 1 proc per session.
(7 + 2*1) * 50,000 = 450,000 procs
RabbitMQ 3.x:
```
erlang:memory().
[{total,15168232704},
{processes,14044779256},
{processes_used,14044755120},
{system,1123453448},
{atom,1057033},
{atom_used,1052587},
{binary,236381264},
{code,21790238},
{ets,391423744}]
erlang:system_info(process_count).
1850309
```
7 procs per connection + 15 per session
(7 + 2*15) * 50,000 = 1,850,000 procs
50k connections + 100k session require
with this commit: 4.5 GB
in RabbitMQ 3.x: 15 GB
## Future work
1. More efficient parser and serializer
2. TODO in mc_amqp: Do not store the parsed message on disk.
3. Implement both AMQP HTTP extension and AMQP management extension to allow AMQP
clients to create RabbitMQ objects (queues, exchanges, ...).
[Why]
This work started as an effort to add peer discovery support to our
Khepri integration. Indeed, as part of the task to integrate Khepri, we
missed the fact that `rabbit_peer_discovery:maybe_create_cluster/1` was
called from the Mnesia-specific code only. Even though we knew about it
because we hit many issues caused by the fact the `join_cluster` and
peer discovery use different code path to create a cluster.
To add support for Khepri, the first version of this patch was to move
the call to `rabbit_peer_discovery:maybe_create_cluster/1` from
`rabbit_db_cluster` instead of `rabbit_mnesia`. To achieve that, it made
sense to unify the code and simply call `rabbit_db_cluster:join/2`
instead of duplicating the work.
Unfortunately, doing so highlighted another issue: the way the node to
cluster with was selected. Indeed, it could cause situations where
multiple clusters are created instead of one, without resorting to
out-of-band counter-measures, like a 30-second delay added in the
Kubernetes operator (rabbitmq/cluster-operator#1156). This problem was
even more frequent when we tried to unify the code path and call
`join_cluster`.
After several iterations on the patch and even more discussions with the
team, we decided to rewrite the algorithm to make node selection more
robust and still use `rabbit_db_cluster:join/2` to create the cluster.
[How]
This commit is only about the rewrite of the algorithm. Calling peer
discovery from `rabbit_db_cluster` instead of `rabbit_mnesia` (and thus
making peer discovery work with Khepri) will be done in a follow-up
commit.
We wanted the new algorithm to fulfill the following properties:
1. `rabbit_peer_discovery` should provide the ability to re-trigger it
easily to re-evaluate the cluster. The new public API is
`rabbit_peer_discovery:sync_desired_cluster/0`.
2. The selection of the node to join should be designed in a way that
all nodes select the same, regardless of the order in which they
become available. The adopted solution is to sort the list of
discovered nodes with the following criterias (in that order):
1. the size of the cluster a discovered node is part of; sorted from
bigger to smaller clusters
2. the start time of a discovered node; sorted from older to younger
nodes
3. the name of a discovered node; sorted alphabetically
The first node in that list will not join anyone and simply proceed
with its boot process. Other nodes will try to join the first node.
3. To reduce the chance of incorrectly having multiple standalone nodes
because the discovery backend returned only a single node, we want to
apply the following constraints to the list of nodes after it is
filtered and sorted (see property 2 above):
* The list must contain `node()` (i.e. the node running peer
discovery itself).
* If the RabbitMQ's cluster size hint is greater than 1, the list
must have at least two nodes. The cluster size hint is the maximum
between the configured target cluster size hint and the number of
elements in the nodes list returned by the backend.
If one of the constraint is not met, the entire peer discovery
process is restarted after a delay.
4. The lock is acquired only to protect the actual join, not the
discovery step where the backend is queried to get the list of peers.
With the node selection described above, this will let the first node
to start without acquiring the lock.
5. The cluster membership views queried as part of the algorithm to sort
the list of nodes will be used to detect additional clusters or
standalone nodes that did not cluster correctly. These nodes will be
asked to re-evaluate peer discovery to increase the chance of forming
a single cluster.
6. After some delay, peer discovery will be re-evaluated to further
eliminate the chances of having multiple clusters instead of one.
This commit covers properties from point 1 to point 4. Remaining
properties will be the scope of additional pull requests after this one
works.
If there is a failure at any point during discovery, filtering/sorting,
locking or joining, the entire process is restarted after a delay. This
is configured using the following parameters:
* cluster_formation.discovery_retry_limit
* cluster_formation.discovery_retry_interval
The default parameters were bumped to 30 retries with a delay of 1
second between each.
The locking retries/interval parameters are not used by the new
algorithm anymore.
There are extra minor changes that come with the rewrite:
* The configured backend is cached in a persistent term. The goal is to
make sure we use the same backend throughout the entire process and
when we call `maybe_unregister/0` even if the configuration changed
for whatever reason in between.
* `maybe_register/0` is called from `rabbit_db_cluster` instead of at
the end of a successful peer discovery process. `rabbit_db_cluster`
had to call `maybe_register/0` if the node was not virgin anyway. So
make it simpler and always call it in `rabbit_db_cluster` regardless
of the state of the node.
* `log_configured_backend/0` is gone. `maybe_init/0` can log the backend
directly. There is no need to explicitly call another function for
that.
* Messages are logged using `?LOG_*()` macros instead of the old
`rabbit_log` module.
[Why]
When a Khepri-based node joins a Mnesia-based cluster, it is reset and
switches back from Khepri to Mnesia. If there are Mnesia files left in
its data directory, Mnesia will restart with stale/incorrect data and
the operation will fail.
After a migration to Khepri, we need to make sure there is no stale
Mnesia files.
[How]
We use `rabbit_mnesia` to query the Mnesia files and delete them.
Providing a pre-hashed and salted password is
not significantly more secure but satisfies those
who cannot pass clear text passwords on the command
line for regulatory reasons.
Note that the optimal way of seeding users is still
definition import on node boot, not scripting with
CLI tools.
Closes#9166
Because both `add_member` and `grow` default to Membership status `promotable`,
new members will have to catch up before they are considered cluster members.
This can be overridden with either `voter` or (permanent `non_voter` statuses.
The latter one is useless without additional tooling so kept undocumented.
- non-voters do not affect quorum size for election purposes
- `observer_cli` reports their status with lowercase 'f'
- `rabbitmq-queues check_if_node_is_quorum_critical` takes voter status into
account
[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>
Michael Klishin
Michael Davis
Jean-Sébastien Pédron
Diana Parra Corbacho
Loïc Hoguin
Michael Klishin
Michal Kuratczyk
Michael Klishin
Rin Kuryloski
Diana Parra Corbacho
Michael Klishin
Ayanda Dube
David Ansari
Alex Valiushko