Jump to: navigation, search


< Zaqar
Revision as of 15:12, 7 November 2014 by Vkmc (talk | contribs) (Marconi API v2.0 - Ideas for review)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)

Marconi API v2.0



  • Add a Swift storage backend
  • Remove SQLAlchemy for realz
  • Remove get-message-by-id endopoint
  • Rename Zaqar to Zarconi (jk)


  • Consider removing include_claimed and always listing all messages
    • Doesn't seem to be terribly useful, since when observing a task distribution queue, you are likely auditing or diagnosing, and will want to see all messages?
  • Remove getting one or more messages by ID
    • Isn't terribly useful to users, and may (TBD) provide for performance optimizations in the drivers
  • Add support for STOMP and/or MQTT?
    • Or is this scope creep? Would it be better to leave it to other messaging systems to deliver on the use cases that these protocols enable?
  • Migrate to topic-based semantics.
    • Can post a single message to multiple topics in one request
    • Can consumers read more than one topic in a single request?
  • Delete a claim to delete all it's messages?
    • Need to decide if we want to encourage deleting multiple messages at a time, or if that is an anti-pattern
  • Only provide a message ID for claimed messages?
    • Con: user may want to have an auditor that uses the feeds api, and the auditor will need to match up message IDs with the ones being claimed by the workers
  • Support msgpack in addition to JSON?
    • Main difficulty is in replying in JSON when a message was enqueued using msgpack, and includes one or more binary fields.
    • Check out Protobuf and Apache Arvo
  • Allow custom media types?
    • Message attributes move into X-headers
    • Allows for all kinds of binary types
    • Con: if an error occurs, there isn't an obvious media type the client accepts to use in reporting structured error information
    • Con: X-headers are unwieldy if we need structured message attributes (YAGNI?)
  • Things that may allow the use for a wider variety of backends [/flameon]
    • For example, make paging through messages without claiming them optional
    • But if people just want Kafka or Qpid or whatever, why not just use those technologies directly? Once you put a REST API on top, it starts to look a lot like v1.x. For extremely high-throughput use cases, apps won't ever get what they want over python + REST. But, on the other hand, there are important use cases where REST is preferable, and if we continue to hold at ~20ms response times, that is still pretty darn good performance for going over the web.