Uploaded image for project: 'Core Server'
  1. Core Server
  2. SERVER-12662

Two background index builds in parallel with unique key constraint violations crashes secondary



    • Icon: Bug Bug
    • Resolution: Done
    • Icon: Critical - P2 Critical - P2
    • 2.4.10
    • 2.4.9
    • Index Maintenance
    • None
    • ALL


      Issue Status as of March 27, 2014

      If a background index build is already in progress, issuing the same index creation command will report success immediately. While this in itself is not a problem, if replication is active, the index creation command will be replicated causing possible problems on secondaries. This will become a problem if the index build ultimately has an error, due to a unique key constraint violation for example. Because of how replication works this can lead to secondary nodes shutting down and not recovering on a restart, since this is an unrecoverable data replication error which causes the server to to shutdown (Fatal Assertion 16360, exception during application of oplog operations on secondaries) for data integrity reasons.

      Users are affected if they create the same background index multiple times concurrently, for example in environments with automated index creation. Due to the fact that the secondaries will shut down during an unrecoverable replication error (to protect data integrity and consistency), this can cause loss of majority and a primary step-down, rendering the replica set read-only.

      The fix is to only replicate index creation commands if they successfully build an index, and not in the case the index already exists. This will prevent secondaries from attempting an index build that could fail, and wasn't successful on the primary.

      To avoid the issue, ensure that background indexes creation commands are only issued one at a time, for each index definition.

      All versions from 2.4.0 to 2.4.9 are affected. The 2.2 and 2.6 production release cycles are unaffected by this issue.

      The fix is included in the production release version 2.4.10.

      Original Description

      If you attempt to build an index that is already in the process of being built (e.g. a background index), the attempt succeeds immediately and also is replicated immediately, even if the in-progress index build might eventually fail (due to, say, a unique constraint violation). This could result in SECONDARY nodes in a replica set attempting index builds that are destined to fail. Such SECONDARY nodes will shut down when the index build fails, since this is a data replication error that cannot be automatically resolved by the server.

      The code fix was to not replicate index builds in the case that such an index already exists.
      Original description and example reproducer:

      On two-member replica set prepare collection like this:

      function prepare() {
          for (var i=0; i<10000; i++)

      Then start two threads in parallel each doing the following index creations:

      function index() {
          while (true) {
              db.c.ensureIndex({x:1}, {unique:true, background:true})

      (I think the while loop isn't necessary as long as the two threads start at close enough to the same time.)

      Index build fails as expected on primary:

      Sun Feb  9 10:46:53.380 [conn7] background addExistingToIndex exception E11000 duplicate key error index: test.c.$x_1  dup key: { : null }
      Sun Feb  9 10:46:53.380 [conn7] end connection (0 connections now open)

      But secondary fasserts:

      Sun Feb  9 10:46:52.164 [repl writer worker 1] info: indexing in foreground on this replica; was a background index build on the primary
      Sun Feb  9 10:46:52.164 [repl writer worker 1] build index test.c { x: 1.0 }
      Sun Feb  9 10:46:52.182 [repl writer worker 1] ERROR: writer worker caught exception: E11000 duplicate key error index: test.c.$x_1  dup key: { : null } on: { ts: Timestamp 1391960812000|1778, h: 5159704\
      80723141670, v: 2, op: "i", ns: "test.system.indexes", o: { _id: ObjectId('52f7a2ec7ddfdca0480c6170'), ns: "test.c", key: { x: 1.0 }, name: "x_1", unique: true, background: true } }
      Sun Feb  9 10:46:52.183 [repl writer worker 1]   Fatal Assertion 16360
      0xde46e1 0xda44d3 0xc2ca7c 0xdb2021 0xe2cf69 0x30040079d1 0x30014e8b6d
       mongod(_ZN5mongo15printStackTraceERSo+0x21) [0xde46e1]
       mongod(_ZN5mongo13fassertFailedEi+0xa3) [0xda44d3]
       mongod(_ZN5mongo7replset14multiSyncApplyERKSt6vectorINS_7BSONObjESaIS2_EEPNS0_8SyncTailE+0x12c) [0xc2ca7c]
       mongod(_ZN5mongo10threadpool6Worker4loopEv+0x281) [0xdb2021]
       mongod() [0xe2cf69]
       /lib64/libpthread.so.0() [0x30040079d1]
       /lib64/libc.so.6(clone+0x6d) [0x30014e8b6d]
      Sun Feb  9 10:46:52.184 [repl writer worker 1]

      It appears that an index creation oplog entry gets made in this case in spite of the index creation failure.




            eliot Eliot Horowitz (Inactive)
            bruce.lucas@mongodb.com Bruce Lucas (Inactive)
            2 Vote for this issue
            9 Start watching this issue