Skip to main content Link Menu Expand (external link) Document Search Copy Copied

High Availability in Client Libraries

Introduction

This article introduces readers to the notion of high availability (“HA”) in BlazingMQ client libraries (also referred to as SDKs). We will see how high availability in client libraries helps users write simpler applications than before, while protecting applications from transient issues in the BlazingMQ back-end, network issues, etc.

High availability in client libraries complements High Availability in BlazingMQ Back-end, and they work together to provide a seamless experience to BlazingMQ applications in case of framework crashes, machine issue, network faults, etc.


Motivation

Looking at BlazingMQ’s network topology, one can see that there are several points of failure along the producer -> primary node as well as primary node -> consumer paths:

  1. Links between proxy and cluster nodes can go down

  2. Links among cluster nodes can go down

  3. Any number of nodes in the cluster can crash

  4. Links between applications and proxies can go down (as a result of proxy crash etc).

Any of these events disrupt data flow and can impact applications. As a result of HA in BlazingMQ back-end, applications are protected from events in 1-3 (readers can refer to HA in BlazingMQ back-end article for more details).

However, event (4) can still impact applications. If BlazingMQ node to which applications are connected to (typically referred as “BlazingMQ proxy” or “local BlazingMQ proxy”) goes down, applications can notice this behavior:

  1. Producer application will get an error when attempting to post a PUT message

  2. Consumer application will get an error which attempting to confirm a message

  3. Producer or consumer application will get an error when attempting to:

    • open a new queue
    • configure an existing queue
    • close a queue

Any of these events can be very disruptive for applications and they need to implement a non-trivial state machine to handle scenarios where new work needs to be submitted to BlazingMQ, but SDK is not connected to BlazingMQ proxy. Lets take the example of a producer application and see how it gets affected:

  • A producer application needs to start buffering any PUT messages while connection with BlazingMQ proxy is down, and then transmit them once connectivity with BlazingMQ is reestablished.

  • Additionally, status of pending PUT messages (messages for which ACKs were not received before connection with BlazingMQ went down) becomes unknown. They may or may not have reached the BlazingMQ queue. Due to this ambiguity, prior to HA work, BlazingMQ SDK would generate negative ACKs for such PUT messages with status = UNKNOWN, indicating to the producer application that the message may or may not have made it to the queue. Some applications could choose to retransmit such PUT messages upon reconnectivity, with the possibility that same message could now appear in the queue twice. It is worth noting that the two copies of this message would have different Message GUIDs assigned to them, making it further challenging for consumer applications to deduplicate such copies.


Design

Just like high availability in BlazingMQ back-end, all cases described above can be solved by buffering any work submitted to the SDK when connectivity is down, and transmitting buffered work and retransmitting any pending work upon reconnection. HA in BlazingMQ SDK takes the same approach. Specifically:

  1. All PUT messages which are submitted by producer application are buffered by the SDK in a collection, irrespective of the state of connectivity with BlazingMQ back-end. This means that once the queue has been successfully opened by the producer application, it can continue to post PUT messages without worrying about connection’s status. This is one of the most important changes in the behavior of the SDK. Previously, producer application would get a NOT_CONNECTED error code from the API when attempting to post PUT message during disconnection, and would need to buffer such PUT messages. This is no longer required.

  2. In addition, SDK now generates and assigns a unique bmqt::MessageGUID to every PUT message submitted by the application. Prior to HA in SDK, GUIDs were assigned by the first BlazingMQ back-end (hop closest to the producer application, typically the local BlazingMQ proxy). Motivation for this change will be explained shortly.

  3. Upon receiving ACK message from BlazingMQ, SDK removes the corresponding PUT message from the collection mentioned in (1) above, since that PUT message is no longer pending (unacknowledged).

  4. When connection with BlazingMQ proxy is down, any new PUT messages submitted by the application will continue to be buffered in the collection in (1).

  5. Once the connection is restored and all queues are reopened, all PUT messages in the collection are transmitted to BlazingMQ back-end. It is important to note that some of these PUT message could have been sent to BlazingMQ back-end prior to connection drop and SDK was waiting for ACKs for them. Moreover, some of these PUT messages could have been accepted by BlazingMQ queue. Such PUT messages will be seamlessly deduped by BlazingMQ back-end, because BlazingMQ primary node maintains a history of MessageGUIDs seen in the last few minutes (configurable), and if a PUT message arrives on the queue with a GUID which exists in the historical list of GUIDs, it is simply acknowledged back without being added to the queue. This ensures that producer application does not see any error, and only one copy of the message appears in the queue. For this logic to work correctly, it is important that the source (SDK) itself generates and assigns a GUID to every PUT message, as discussed in (2) above.

  6. Unlike PUT messages, CONFIRM messages will not be buffered by the SDK if connection to BlazingMQ proxy is down. While this may seem inconsistent, there is a good reason for this behavior – very often, there are multiple consumers attached to a queue. If a BlazingMQ node loses route to the consumer which is processing PUSH messages and sending CONFIRMs, it is extremely likely that the BlazingMQ node will pick another consumer on a different route and start delivering those unconfirmed PUSH messages to the new consumer immediately. As such, even if original consumer buffers CONFIRM messages when it is disconnected to BlazingMQ proxy and transmits them upon reconnection, those CONFIRM messages will simply be ignored by BlazingMQ primary node as they are stale, and those PUSH messages will get processed twice – once by original consumer and then by new consumer. Note that this behavior is within contract, as BlazingMQ provides at least once delivery guarantee by design. We do have some ideas to minimize such duplicate PUSH messages, and we may work on it in the near future.

  7. OpenQueue and ConfigureQueue requests will be buffered during disconnection and transmitted upon reconnection. Additionally, if a request is pending while connection goes down, it will not be failed, but instead seamlessly resent upon reconnection.

  8. Lastly, CloseQueue request is not buffered during disconnection. Queue is simply marked as closed in the SDK and a response is generated locally. Upon reconnection, this queue is not reopened, effectively leaving the queue in the closed state.

TCP High Watermark Handling

Prior to HA logic in SDK, if TCP connection between SDK and BlazingMQ proxy is flow-controlled such that SDK receives a push back from TCP socket when attempting to send something, the SDK would buffer messages up to a certain limit (typically 128MB), and any attempt to send messages beyond that limit would return BW_LIMIT (Bandwidth Limit) error to the application, and in some cases, would simply drop the TCP connection.

This is no longer the case. SDK will continue to accept messages from the application even when it is getting a push back from TCP socket, and instead of enforcing a size limit, it now enforces a time limit of 5 seconds (non-configurable, hard-coded in the SDK) for the TCP channel to be “open” again. After this time interval, if TCP channel is still flow-controlled, SDK will return BW_LIMIT error to the application. It is important to note that the wait for this time interval occurs in the application thread which calls the SDK API to send message. This is done to ensure that the push back indicated by TCP layer is exposed all the way back to the application.