diff options
author | Alan Conway <aconway@apache.org> | 2012-02-13 23:50:18 +0000 |
---|---|---|
committer | Alan Conway <aconway@apache.org> | 2012-02-13 23:50:18 +0000 |
commit | 057a0635e081848ba59964c4a8e0923e521b7fe6 (patch) | |
tree | cdad9578140d5dbdb2e90525e2ba9cc870ca50b6 /qpid/cpp/design_docs/new-ha-design.txt | |
parent | cf61dbf9b313f9bd69b392eae8fd8d27d4e609a2 (diff) | |
download | qpid-python-057a0635e081848ba59964c4a8e0923e521b7fe6.tar.gz |
Merge branch 'qpid-3603-4-rebase' into qpid-3603-5qpid-3603-5
git-svn-id: https://svn.apache.org/repos/asf/qpid/branches/qpid-3603-5@1243748 13f79535-47bb-0310-9956-ffa450edef68
Diffstat (limited to 'qpid/cpp/design_docs/new-ha-design.txt')
-rw-r--r-- | qpid/cpp/design_docs/new-ha-design.txt | 418 |
1 files changed, 418 insertions, 0 deletions
diff --git a/qpid/cpp/design_docs/new-ha-design.txt b/qpid/cpp/design_docs/new-ha-design.txt new file mode 100644 index 0000000000..8173585283 --- /dev/null +++ b/qpid/cpp/design_docs/new-ha-design.txt @@ -0,0 +1,418 @@ +-*-org-*- +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. + +* An active-passive, hot-standby design for Qpid clustering. + +This document describes an active-passive approach to HA based on +queue browsing to replicate message data. + +See [[./old-cluster-issues.txt]] for issues with the old design. + +** Active-active vs. active-passive (hot-standby) + +An active-active cluster allows clients to connect to any broker in +the cluster. If a broker fails, clients can fail-over to any other +live broker. + +A hot-standby cluster has only one active broker at a time (the +"primary") and one or more brokers on standby (the "backups"). Clients +are only served by the primary, clients that connect to a backup are +redirected to the primary. The backups are kept up-to-date in real +time by the primary, if the primary fails a backup is elected to be +the new primary. + +The main problem with active-active is co-ordinating consumers of the +same queue on multiple brokers such that there are no duplicates in +normal operation. There are 2 approaches: + +Predictive: each broker predicts which messages others will take. This +the main weakness of the old design so not appealing. + +Locking: brokers "lock" a queue in order to take messages. This is +complex to implement and it is not straighforward to determine the +best strategy for passing the lock. In tests to date it results in +very high latencies (10x standalone broker). + +Hot-standby removes this problem. Only the primary can modify queues +so it just has to tell the backups what it is doing, there's no +locking. + +The primary can enqueue messages and replicate asynchronously - +exactly like the store does, but it "writes" to the replicas over the +network rather than writing to disk. + +** Replicating browsers + +The unit of replication is a replicating browser. This is an AMQP +consumer that browses a remote queue via a federation link and +maintains a local replica of the queue. As well as browsing the remote +messages as they are added the browser receives dequeue notifications +when they are dequeued remotely. + +On the primary broker incoming mesage transfers are completed only when +all of the replicating browsers have signaled completion. Thus a completed +message is guaranteed to be on the backups. + +** Failover and Cluster Resource Managers + +We want to delegate the failover management to an existing cluster +resource manager. Initially this is rgmanager from Cluster Suite, but +other managers (e.g. PaceMaker) could be supported in future. + +Rgmanager takes care of starting and stopping brokers and informing +brokers of their roles as primary or backup. It ensures there's +exactly one primary broker running at any time. It also tracks quorum +and protects against split-brain. + +Rgmanger can also manage a virtual IP address so clients can just +retry on a single address to fail over. Alternatively we will also +support configuring a fixed list of broker addresses when qpid is run +outside of a resource manager. + +Aside: Cold-standby is also possible using rgmanager with shared +storage for the message store (e.g. GFS). If the broker fails, another +broker is started on a different node and and recovers from the +store. This bears investigation but the store recovery times are +likely too long for failover. + +** Replicating configuration + +New queues and exchanges and their bindings also need to be replicated. +This is done by a QMF client that registers for configuration changes +on the remote broker and mirrors them in the local broker. + +** Use of CPG (openais/corosync) + +CPG is not required in this model, an external cluster resource +manager takes care of membership and quorum. + +** Selective replication + +In this model it's easy to support selective replication of individual queues via +configuration. + +Explicit exchange/queue qpid.replicate argument: +- none: the object is not replicated +- configuration: queues, exchanges and bindings are replicated but messages are not. +- messages: configuration and messages are replicated. + +TODO: provide configurable default for qpid.replicate + +[GRS: current prototype relies on queue sequence for message identity +so selectively replicating certain messages on a given queue would be +challenging. Selectively replicating certain queues however is trivial.] + +** Inconsistent errors + +The new design eliminates most sources of inconsistent errors in the +old design (connections, sessions, security, management etc.) and +eliminates the need to stall the whole cluster till an error is +resolved. We still have to handle inconsistent store errors when store +and cluster are used together. + +We have 2 options (configurable) for handling inconsistent errors, +on the backup that fails to store a message from primary we can: +- Abort the backup broker allowing it to be re-started. +- Raise a critical error on the backup broker but carry on with the message lost. +We can configure the option to abort or carry on per-queue, we +will also provide a broker-wide configurable default. + +** New backups connecting to primary. + +When the primary fails, one of the backups becomes primary and the +others connect to the new primary as backups. + +The backups can take advantage of the messages they already have +backed up, the new primary only needs to replicate new messages. + +To keep the N-way guarantee the primary needs to delay completion on +new messages until all the back-ups have caught up. However if a +backup does not catch up within some timeout, it is considered dead +and its messages are completed so the cluster can carry on with N-1 +members. + + +** Broker discovery and lifecycle. + +The cluster has a client URL that can contain a single virtual IP +address or a list of real IP addresses for the cluster. + +In backup mode, brokers reject connections normal client connections +so clients will fail over to the primary. HA admin tools mark their +connections so they are allowed to connect to backup brokers. + +Clients discover the primary by re-trying connection to the client URL +until the successfully connect to the primary. In the case of a +virtual IP they re-try the same address until it is relocated to the +primary. In the case of a list of IPs the client tries each in +turn. Clients do multiple retries over a configured period of time +before giving up. + +Backup brokers discover the primary in the same way as clients. There +is a separate broker URL for brokers since they often will connect +over a different network. The broker URL has to be a list of real +addresses rather than a virtual address. + +Brokers have the following states: +- connecting: Backup broker trying to connect to primary - loops retrying broker URL. +- catchup: Backup connected to primary, catching up on pre-existing configuration & messages. +- ready: Backup fully caught-up, ready to take over as primary. +- primary: Acting as primary, serving clients. + +** Interaction with rgmanager + +rgmanager interacts with qpid via 2 service scripts: backup & +primary. These scripts interact with the underlying qpidd +service. rgmanager picks the new primary when the old primary +fails. In a partition it also takes care of killing inquorate brokers. + +*** Initial cluster start + +rgmanager starts the backup service on all nodes and the primary service on one node. + +On the backup nodes qpidd is in the connecting state. The primary node goes into +the primary state. Backups discover the primary, connect and catch up. + +*** Failover + +primary broker or node fails. Backup brokers see disconnect and go +back to connecting mode. + +rgmanager notices the failure and starts the primary service on a new node. +This tells qpidd to go to primary mode. Backups re-connect and catch up. + +The primary can only be started on nodes where there is a ready backup service. +If the backup is catching up, it's not eligible to take over as primary. + +*** Failback + +Cluster of N brokers has suffered a failure, only N-1 brokers +remain. We want to start a new broker (possibly on a new node) to +restore redundancy. + +If the new broker has a new IP address, the sysadmin pushes a new URL +to all the existing brokers. + +The new broker starts in connecting mode. It discovers the primary, +connects and catches up. + +*** Failure during failback + +A second failure occurs before the new backup B completes its catch +up. The backup B refuses to become primary by failing the primary +start script if rgmanager chooses it, so rgmanager will try another +(hopefully caught-up) backup to become primary. + +*** Backup failure + +If a backup fails it is re-started. It connects and catches up from scratch +to become a ready backup. + +** Interaction with the store. + +Clean shutdown: entire cluster is shut down cleanly by an admin tool: +- primary stops accepting client connections till shutdown is complete. +- backups come fully up to speed with primary state. +- all shut down marking stores as 'clean' with an identifying UUID. + +After clean shutdown the cluster can re-start automatically as all nodes +have equivalent stores. Stores starting up with the wrong UUID will fail. + +Stored status: clean(UUID)/dirty, primary/backup, generation number. +- All stores are marked dirty except after a clean shutdown. +- Generation number: passed to backups and incremented by new primary. + +After total crash must manually identify the "best" store, provide admin tool. +Best = highest generation number among stores in primary state. + +Recovering from total crash: all brokers will refuse to start as all stores are dirty. +Check the stores manually to find the best one, then either: +1. Copy stores: + - copy good store to all hosts + - restart qpidd on all hosts. +2. Erase stores: + - Erase the store on all other hosts. + - Restart qpidd on the good store and wait for it to become primary. + - Restart qpidd on all other hosts. + +Broker startup with store: +- Dirty: refuse to start +- Clean: + - Start and load from store. + - When connecting as backup, check UUID matches primary, shut down if not. +- Empty: start ok, no UUID check with primary. + +** Current Limitations + +(In no particular order at present) + +For message replication: + +LM1 - The re-synchronisation does not handle the case where a newly elected +primary is *behind* one of the other backups. To address this I propose +a new event for restting the sequence that the new primary would send +out on detecting that a replicating browser is ahead of it, requesting +that the replica revert back to a particular sequence number. The +replica on receiving this event would then discard (i.e. dequeue) all +the messages ahead of that sequence number and reset the counter to +correctly sequence any subsequently delivered messages. + +LM2 - There is a need to handle wrap-around of the message sequence to avoid +confusing the resynchronisation where a replica has been disconnected +for a long time, sufficient for the sequence numbering to wrap around. + +LM3 - Transactional changes to queue state are not replicated atomically. + +LM4 - Acknowledgements are confirmed to clients before the message has been +dequeued from replicas or indeed from the local store if that is +asynchronous. + +LM5 - During failover, messages (re)published to a queue before there are +the requisite number of replication subscriptions established will be +confirmed to the publisher before they are replicated, leaving them +vulnerable to a loss of the new primary before they are replicated. + +LM6 - persistence: In the event of a total cluster failure there are +no tools to automatically identify the "latest" store. Once this +is manually identfied, all other stores belonging to cluster members +must be erased and the latest node must started as primary. + +LM6 - persistence: In the event of a single node failure, that nodes +store must be erased before it can re-join the cluster. + +For configuration propagation: + +LC2 - Queue and exchange propagation is entirely asynchronous. There +are three cases to consider here for queue creation: + +(a) where queues are created through the addressing syntax supported +the messaging API, they should be recreated if needed on failover and +message replication if required is dealt with seperately; + +(b) where queues are created using configuration tools by an +administrator or by a script they can query the backups to verify the +config has propagated and commands can be re-run if there is a failure +before that; + +(c) where applications have more complex programs on which +queues/exchanges are created using QMF or directly via 0-10 APIs, the +completion of the command will not guarantee that the command has been +carried out on other nodes. + +I.e. case (a) doesn't require anything (apart from LM5 in some cases), +case (b) can be addressed in a simple manner through tooling but case +(c) would require changes to the broker to allow client to simply +determine when the command has fully propagated. + +LC3 - Queues that are not in the query response received when a +replica establishes a propagation subscription but exist locally are +not deleted. I.e. Deletion of queues/exchanges while a replica is not +connected will not be propagated. Solution is to delete any queues +marked for propagation that exist locally but do not show up in the +query response. + +LC4 - It is possible on failover that the new primary did not +previously receive a given QMF event while a backup did (sort of an +analogous situation to LM1 but without an easy way to detect or remedy +it). + +LC5 - Need richer control over which queues/exchanges are propagated, and +which are not. + +LC6 - The events and query responses are not fully synchronized. + + In particular it *is* possible to not receive a delete event but + for the deleted object to still show up in the query response + (meaning the deletion is 'lost' to the update). + + It is also possible for an create event to be received as well + as the created object being in the query response. Likewise it + is possible to receive a delete event and a query response in + which the object no longer appears. In these cases the event is + essentially redundant. + + It is not possible to miss a create event and yet not to have + the object in question in the query response however. + + +* Benefits compared to previous cluster implementation. + +- Does not depend on openais/corosync, does not require multicast. +- Can be integrated with different resource managers: for example rgmanager, PaceMaker, Veritas. +- Can be ported to/implemented in other environments: e.g. Java, Windows +- Disaster Recovery is just another backup, no need for separate queue replication mechanism. +- Can take advantage of resource manager features, e.g. virtual IP addresses. +- Fewer inconsistent errors (store failures) that can be handled without killing brokers. +- Improved performance +* User Documentation Notes + +Notes to seed initial user documentation. Loosely tracking the implementation, +some points mentioned in the doc may not be implemented yet. + +** High Availability Overview + +HA is implemented using a 'hot standby' approach. Clients are directed +to a single "primary" broker. The primary executes client requests and +also replicates them to one or more "backup" brokers. If the primary +fails, one of the backups takes over the role of primary carrying on +from where the primary left off. Clients will fail over to the new +primary automatically and continue their work. + +TODO: at least once, deduplication. + +** Enabling replication on the client. + +To enable replication set the qpid.replicate argument when creating a +queue or exchange. + +This can have one of 3 values +- none: the object is not replicated +- configuration: queues, exchanges and bindings are replicated but messages are not. +- messages: configuration and messages are replicated. + +TODO: examples +TODO: more options for default value of qpid.replicate + +A HA client connection has multiple addresses, one for each broker. If +the it fails to connect to an address, or the connection breaks, +it will automatically fail-over to another address. + +Only the primary broker accepts connections, the backup brokers +redirect connection attempts to the primary. If the primary fails, one +of the backups is promoted to primary and clients fail-over to the new +primary. + +TODO: using multiple-address connections, examples c++, python, java. + +TODO: dynamic cluster addressing? + +TODO: need de-duplication. + +** Enabling replication on the broker. + +Network topology: backup links, separate client/broker networks. +Describe failover mechanisms. +- Client view: URLs, failover, exclusion & discovery. +- Broker view: similar. +Role of rmganager + +** Configuring rgmanager + +** Configuring qpidd + + |