Apache Kafka integration

From Fusion Registry Wiki
Revision as of 09:36, 30 November 2020 by Mnelson (talk | contribs) (Publication Reliability)
Jump to navigation Jump to search

Compatibility

Product Module Version Support
Fusion Registry Enterprise Edition Core 10.0 Kafka producer supporting 'Structure Notification' events

Overview

Fusion Registry can act as an Apache Kafka Producer where specified events are published on definable Kafka topics. It consists of a generalised ‘producer’ interface capable of publishing any information to definable Kafka topics, and a collection of ‘handlers’ for managing specific events.

At present there is only one event handler Structure Notification which publishes changes to any structures as they occur, this also include data registrations.

For changes or modifications to structures, the body of the message is always an SDMX structure message. The format is configurable at the event handler level, the structure producer can support any structure output as long as the registry is aware of the VND header, however the UI offers the following choices:

  • Fusion JSON (the JSON dialect that pre-dated the formal SDMX-JSON specification)
  • SDMX-ML v2.1

Fusion JSON is recommended if the Fusion Registry is being used to support non-sdmx structures (which can not be described in SDMX-ML), for example Reporting Templates and Validation Schemes.

A ‘tombstone’ message is used for structure deletions with the structure URN in the Kafka Message Key, and a null body.

The list of supported events that can be published to Kafka will be extended over time.

Other events envisaged include:

  • Audit events such as logins, data registrations and API calls
  • Errors, for instance where a scheduled data import fails
  • Configuration changes
  • Changes to Content Security Rules

Configuration

Configuration is performed through the GUI with 'admin' privileges.

Connection

Before Kafka can be used, a connection must be defined the Fusion Registry to a Kafka server (or servers), taking note of the following.

Number of Connections

The Registry supports one connection definition to one or more bootstrap servers. Each server connection can be separated by a comma. The UI displays all connection properties supported by Kafka, and displays the default value that will be used when not explicitly set.

Transactions

Whilst Kafka does not enforce transactions, the Fusion Registry does. A connection will default to using the transaction prefix ‘tx’. The documentation states that this can be null, which is not the case for the Fusion Registry.

Kerberos

showing settings to support kerberos authentication

Kerberos integration can be achieved by providing a value for the sasl.jaas.config property. This property expects a new line per configuration setting, in the UI the text area is provided to support multiple lines.

Max Request Size

The connection definition supports the max.request.size property, which will enforce a restriction on the maximum size of the message that can be sent. Additional restrictions for this can be defined on the Kafka server, and the Kafka topic, which are both configured outside of the Fusion Registry. It is important to ensure the server and topic can also handle large request sizes if this property is changed in the Fusion Registry. It is important to note that the max request size refers to the size of the message BEFORE any compression is applied. This is a behaviour of Kafka and cannot be modified. When the message is sent, both the topic and connection will apply the max request size on the compressed message (if compression is enabled) and therefore these can be set to a lower threshold.

Test Connection

Example connection test results

The Fusion Registry ensures the bootstrap servers can all be connected to before the connection is saved. However, the test function should be used to ensure the following:

  • A message can be sent using the security defined in the connection (e.g. Kerberos)
  • Transactions are supported by the connection
  • Large messages can be sent with no issue

Note 1: the last test generates a large message (up to the number of bytes specified). As the message contains a lot of repeated text, the compression (if set) will make the final message size very small. If the intention is to also test the message size against the topic and kafka server, then ensure compression is set to none before running the test.

Note 2: The test is run against the applied configuration, remember to apply changes BEFORE running the test.

Structure Producer

Setting Up

Showing the UI for the Structure Producer

The Structure Producer uses the Kafka connection to send all structural changes. However, before the structure producer becomes operational, it must be assigned to the Kafka topic to use, and must be given a message format to notify structures in.

The structure producer will not become active until it is started. The fist time the producer is started it can either:

  1. Start and remain idle until a structure is added/edited/removed
  2. Start and send all current Fusion Registry structures to it's given topic in a single transaction (beware of the max size set on the connection)

Behaviour

When the Fusion Registry restarts the Structure Producer will always cross check the last transaction successfully sent to the topic against the last transaction the Fusion Registry processed. If the Fusion Registry has transactions not sent to the topic, then the structure producer will go through each unsent transaction, in sequence, until it is up to date.

Note: If the structure producer has never sent a transaction, and the Fusion Registry is restarted, the producer will not replay all transactions, it will remain idle until the next transaction occurs in the Fusion Registry.

The transaction Id shown by the structure producer match those shown in the Fusion Registry RSS feed.

Manual Suspension

The structure producer can be suspended at any time by clicking on the 'suspend' button. Note, that this only suspends the producer for as long as the Fusion Registry is running. On Fusion Registry restart, the producer will always attempt to start if it can. To remove the producer permanently, use the 'delete configuration' option.

Forced Suspension and Error Handling

Kafka is able to detect and handle transitory errors, such as the Kafka server becoming unresponsive for a period of time, there may be occasions where a message simply cannot be sent and no amount of re-attempts will fix this. An example failure is when the message size exceeds the limit set on the connection, server, or topic.

When a failure condition is met by the structure producer which cannot be resolved through re-attempts, the structure producer is shut down, with the transaction state being recorded as failed. It is strongly recommended to configure the Fusion Registry to use an email server, and enter an Admin email address for the Structure Producer. When there is a failure, the Fusion Registry will send an email to this email account, which will include the error message (also found in the Fusion Registry log files).

General Behaviour

Publication Reliability

Message publication has been designed to be 100% reliable, with messages sent in the same sequence that the Fusion Registry processed them.

The Fusion Registry makes use of Kafka supplied libraries to place messages on the topic, any transitory failures are handled by the Kafka library. A transitory failure includes the inability for the Fusion Registry to talk to the kafka server, for example if the kafka server is being restarted. Possible transitory errors in the Fusion Registry are managed by reattempting the send (up to 3 times).

When a message simply cannot be sent, for example the Fusion Registry can no longer authenticate with kafka, or the message size exceeds limits set, the producer is stopped and emails are sent to notify of the failure. Once the problem is rectified and the producer restarted, the Fusion Registry will replay any pending transactions that occurred during the period of time for which the producer was stopped.

Each Fusion Registry transaction is recorded in the database, along with the kafka status. On server restart the Fusion Registry will cross reference the transaction logs with the kafka logs to ensure all messages have been sent to kafka.

The kafka logs are only set to SENT when the Fusion Registry has been notified by the kafka server that it has successfully received the message.

Events

The following Fusion Registry events are available for publication on a Kafka service.

Event Details
Structure Notification Addition / modification or deletion of SDMX structures

Structure Notification

SDMX Structural Metadata is published to definable Kafka topics on Fusion Registry startup and each time structures are added / modified, or deleted. Metadata-driven applications can subscribe to the relevant topic(s) to receive notifications of changes to structures allowing them to maintain an up-to-date replica copy of the structural metadata they need.

Example use cases include:

  • Structural validation applications which require up-to-date structural metadata such as Codelists and DSDs to check whether data is correctly structured.
  • Structure mapping applications which require up-to-date copies of relevant Structure Sets and maps to perform data transformations.


KafkaMetadataNotification.png

Kafka Message Key

The SDMX Structure URN is used as the Kafka Message Key. Consumers must therefore be prepared to receive and interpret the key as the Structure URN correctly when processing the message. This is particularly important for structure deletion where the Message Key is the only source of information about which structure is being referred to.

Kafka Message Body

The Kafka Message Body is a normal SDMX structure message, except in the case of deletions where the message has no content (a null payload), and the structure to delete is identified exclusively by its URN in the Message Key. The format of the SDMX structure message is configurable using the Topics form with the following options available:

Format Explanation HTTP Accept Header Equivalent
SDMX-ML 1.0 SDMX-ML 1.0 XML structure document application/vnd.sdmx.structure+xml;version=1.0
SDMX-ML 2.0 SDMX-ML 2.0 XML structure document application/vnd.sdmx.structure+xml;version=2.0
SDMX-ML 2.0 SDMX-ML 2.1 XML structure document application/vnd.sdmx.structure+xml;version=2.1
SDMX-EDI EDI application/vnd.sdmx.structure;version=edi
SDMX-JSON SDMX-JSON 1.0 application/vnd.sdmx.json
XLSX (flat) Metadata Technology's Excel structure format application/vnd.xlsx
Fusion-JSON Metadata Technology's JSON format with similarities to SDMX-JSON provided primarily for backward compatibility application/vnd.fusion.json
Events for certain structures will only be published if the structure type is supported by the chosen format. Only Fusion-JSON supports all possible structures including Fusion Registry 'extended structures' like Validation Schemes.

Additions and Modifications

Addition and Modification of structures both result in an SDMX 'replace' message containing the full content of the structures. Deltas (such as the addition of a Code to a Codelist) are not supported.

Deletions

Deletion of structures results in a 'tombstone' message, i.e. one with a null payload but the URN of the deleted structure in the Message Key.

Kafka Transactions

Additions and modifications to structures in Fusion Registry are published using Kafka Transactional Messaging to help guarantee consistency by providing:

  • Atomicity - The consumer will not be exposed to uncommited transactions
  • Durability - Kafka brokers guarantee not to lose any committed transactions
  • Ordering - Transaction-aware consumers receive transactions in the original order
  • Non-Duplication - No duplicate messages within transactions

If a single structure is added or changed, the transaction will contain just that structure.

Transactions will contain multiple structures in cases including:

  • Where multiple structures are submitted to the Registry in a single SDMX-ML message.
  • Where a structure's Agency or ID is changed that affects dependent structures.

Subscribers to the Kafka topic must be transaction-aware, and must ensure that each transaction is processed atomically to maintain the referential integrity of their structural metadata replicas.

Structure deletions are never mixed with additions / modifications in the same transaction. This reflects the behaviour of the SDMX REST API whereby additions / modifications use HTTP POST, while deletions require HTTP DELETE.

Full Content Publication and Re-synchronisation

Under certain conditions, the Fusion Registry Producer will publish all of its structures in a single SDMX Structure Message of the chosen format to Kafka.
These are:

  1. When a new Kafka connection is set up, or the configuration of an existing Kafka connection is changed
  2. Registry startup

The principle here is to ensure that Consumers have a consistent baseline against which to process subsequent change notifications.

It also mitigates the risk that changes held in the Registry's in-memory queue of messages awaiting Kafka publication were lost on shutdown by forcing Consumers to re-synchronise on Registry startup.

Not Supported: Subscription to Specific Structures

The current implementation does not allow Kafka Consumers to subscribe to changes on specific structures or sets of structures, such as those maintained by a particular Agency. Consumers subscribing to the chosen topic will receive information about all structures and will need to select what they need.

Not Supported: Publication Arbitration on Fusion Registry Load Balanced Clusters

IMPORTANT: There is currently no arbitration mechanism between peer Fusion Registry instances configured in a Load Balanced cluster. This means that each member of the cluster independently publishes each structure change notification to Kafka, resulting in duplication of notifications.

Structure changes made on one instance of Fusion Registry in a cluster are replicated to all of the others through either a shared database polling mechanism, or by Rabbit MQ messaging. An instance receiving information that structures have changed elsewhere in the cluster updates its in-memory SDMX information model accordingly. The update event is however trapped by its Kafka notification processor and published as normal.

Changes are required to the Kafka event processing behaviour of Registries operating in a cluster to arbitrate over which instance publishes the notification on Kafka such that a structure change is notified once and only once, as expected.

Not Supported: Enforcement of Fusion Registry Content Security Rules to Structures Published on Kafka

Fusion Registry Content Security defines rules restricting access to selected structures and items to specific groups.

The Kafka Structure Notification processor does not enforce Content Security structure rules meaning that all structures are published to the Kafka broker service on the defined topic(s), irrespective of what restrictions may be in place. The risk here is that people or applications that would otherwise not have access to specific structures may be able circumvent the rules normally imposed by Fusion Registry's Content Security sub-system by subscribing to the Structure Notification topic on Kafka. Kafa Streams Security allows control over access to topics. While an all-or-nothing approach of limiting access to all published structures may be sufficient for some applications, it doesn't solve the use case where a consumer should be allowed access to certain structures, but not others.

The Kafka publish-and-subscribe mechanism means that the Fusion Registry producer has no knowledge of who will be consuming the metadata at the point of publication. Up-front decisions therefore have to be made as to what rules to apply when publishing to Kafa. Options include:

  • Only publish 'public' structures - i.e. those without any rules restricting access
  • Publish changes for each Content Security group (and 'public') on their own separate Kafka topics - applying Kafka Security to the topics would then ensure the intended access restrications are enforced