Details

    • Type: New Feature New Feature
    • Status: Open
    • Priority: Major Major
    • Resolution: Unresolved
    • Affects Version/s: 0.8.0
    • Fix Version/s: None
    • Component/s: None
    • Labels:
      None

      Description

      LinkedIn has a system that does monitoring on top of our data flow to ensure all data is delivered to all consumers of data. This works by having each logical "tier" through which data passes produce messages to a central "audit-trail" topic; these messages give a time period and the number of messages that passed through that tier in that time period. Example of tiers for data might be "producer", "broker", "hadoop-etl", etc. This makes it possible to compare the total events for a given time period to ensure that all events that are produced are consumed by all consumers.

      This turns out to be extremely useful. We also have an application that "balances the books" and checks that all data is consumed in a timely fashion. This gives graphs for each topic and shows any data loss and the lag at which the data is consumed (if any).

      This would be an optional feature that would allow you to to this kind of reconciliation automatically for all the topics kafka hosts against all the tiers of applications that interact with the data.

      Some details, the proposed format of the data is JSON using the following format for messages:

      { "time":1301727060032, // the timestamp at which this audit message is sent "topic": "my_topic_name", // the topic this audit data is for "tier":"producer", // a user-defined "tier" name "bucket_start": 1301726400000, // the beginning of the time bucket this data applies to "bucket_end": 1301727000000, // the end of the time bucket this data applies to "host":"my_host_name.datacenter.linkedin.com", // the server that this was sent from "datacenter":"hlx32", // the datacenter this occurred in "application":"newsfeed_service", // a user-defined application name "guid":"51656274-a86a-4dff-b824-8e8e20a6348f", // a unique identifier for this message "count":43634 }

      DISCUSSION

      Time is complex:
      1. The audit data must be based on a timestamp in the events not the time on machine processing the event. Using this timestamp means that all downstream consumers will report audit data on the right time bucket. This means that there must be a timestamp in the event, which we don't currently require. Arguably we should just add a timestamp to the events, but I think it is sufficient for now just to allow the user to provide a function to extract the time from their events.
      2. For counts to reconcile exactly we can only do analysis at a granularity based on the least common multiple of the bucket size used by all tiers. The simplest is just to configure them all to use the same bucket size. We currently use a bucket size of 10 mins, but anything from 1-60 mins is probably reasonable.

      For analysis purposes one tier is designated as the source tier and we do reconciliation against this count (e.g. if another tier has less, that is treated as lost, if another tier has more that is duplication).

      Note that this system makes false positives possible since you can lose an audit message. It also makes false negatives possible since if you lose both normal messages and the associated audit messages it will appear that everything adds up. The later problem is astronomically unlikely to happen exactly, though.

      This would integrate into the client (producer and consumer both) in the following way:
      1. The user provides a way to get timestamps from messages (required)
      2. The user configures the tier name, host name, datacenter name, and application name as part of the consumer and producer config. We can provide reasonable defaults if not supplied (e.g. if it is a Producer then set tier to "producer" and get the hostname from the OS).

      The application that processes this data is currently a Java Jetty app and talks to mysql. It feeds off the audit topic in kafka and runs both automatic monitoring checks and graphical displays of data against this. The data layer is not terribly scalable but because the audit data is sent only periodically this is enough to allow us to audit thousands of servers on very modest hardware, and having sql access makes diving into the data to trace problems to particular hosts easier.

      LOGISTICS
      I would recommend the following steps:
      1. Add the audit application, the proposal would be to add a new top-level directory equivalent to core or perf called "audit" to house this application. At this point it would just be sitting there, not really being used.
      2. Integrate these capabilities into the producer as part of the refactoring we are doing now
      3. Integrate into consumer when possible

      1. kafka-audit-trail-draft.patch
        752 kB
        Jay Kreps
      2. Picture 18.png
        217 kB
        Jay Kreps

        Activity

          People

          • Assignee:
            Jay Kreps
            Reporter:
            Jay Kreps
          • Votes:
            10 Vote for this issue
            Watchers:
            28 Start watching this issue

            Dates

            • Created:
              Updated:

              Development