Uploaded image for project: 'Core Server'
  1. Core Server
  2. SERVER-32448

ShardRegistry::reload() does blocking work on NetworkInterface thread

    XMLWordPrintableJSON

Details

    • Bug
    • Status: Closed
    • Major - P3
    • Resolution: Works as Designed
    • 3.4.14, 3.6.4
    • None
    • Networking
    • Sharding
    • ALL
    • Platforms 2018-01-01, Platforms 2018-01-15

    Description

      The ShardRegistry schedules itself to perform _internalReload()s periodically, using the TaskExecutor. These jobs are dispatched to NetworkInterfaceASIO, and will eventually run on its thread.

      From _internalReload(), we call reload(), which tries to getAllShards() from the ShardingCatalogClientImpl. getAllShards() makes a Fetcher instance, which launches networking work, and then waits for it to join(). However, when we run this, we are already on NetworkInterfaceASIO's thread. This breaks the contract that callbacks to NetworkInterfaceASIO may not perform blocking work. Worse, when these calls are issued through the same TaskExecutor, the thread will deadlock.

      Attachments

        Issue Links

          Activity

            People

              backlog-server-sharding Backlog - Sharding Team
              samantha.ritter@mongodb.com Samantha Ritter (Inactive)
              Votes:
              0 Vote for this issue
              Watchers:
              7 Start watching this issue

              Dates

                Created:
                Updated:
                Resolved: