Uploaded image for project: 'Core Server'
  1. Core Server
  2. SERVER-48566

Shard loader in primary nodes blindly read the version of config.cache.collections

    XMLWordPrintable

    Details

    • Backwards Compatibility:
      Fully Compatible
    • Operating System:
      ALL
    • Backport Requested:
      v4.4
    • Sprint:
      Sharding 2020-06-15, Sharding 2020-06-29
    • Linked BF Score:
      24

      Description

      When the catalog cache loader does a refresh from scratch, it first checks the maximum version that it has from config.cache.collections. This version will be used in the query to get the chunk diffs. It will then combine the result from the diff and config.cache.chunks to form the complete collection metadata. The issue, however, is that the writes in the config.cache.collections and config.cache.chunks are not atomic. This means that it is possible for the version to be higher than what is contained in the config.cache.chunks. This can make the loader falsely believe that it is a certain version, but the contents of the chunk map does not agree with it.

      One example manifestation is that when a shard key refine happens, then the primary stepped down after it modified the config.cache.collections but before it was able to modify the config.cache.chunks (or was rolled back), the new primary will not be able to find the matching exact bounds when trying to migrate because the in memory metadata will still contain the older shard keys despite having the version of post refine.

        Attachments

          Issue Links

            Activity

              People

              Assignee:
              renctan Randolph Tan
              Reporter:
              renctan Randolph Tan
              Participants:
              Votes:
              0 Vote for this issue
              Watchers:
              3 Start watching this issue

                Dates

                Created:
                Updated:
                Resolved: