[HDFS-12862] CacheDirective becomes invalid when NN restart or failover

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.7.1
• Fix Version/s: 3.3.0, 3.2.2
• Component/s: caching, hdfs
• Labels: patch
• Target Version/s: 3.4.0
• Hadoop Flags: Reviewed
• Assignee: Wang XL
• Reporter: Wang XL
• Created: 28/Nov/17 08:48
• Updated: 27/Jan/24 03:32
• Resolved: 12/Apr/20 05:49

Description

The logic in FSNDNCacheOp#modifyCacheDirective is not correct. when modify cacheDirective,the expiration in directive may be a relative expiryTime, and EditLog will serial a relative expiry time.

// Some comments here
static void modifyCacheDirective(
      FSNamesystem fsn, CacheManager cacheManager, CacheDirectiveInfo directive,
      EnumSet<CacheFlag> flags, boolean logRetryCache) throws IOException {
    final FSPermissionChecker pc = getFsPermissionChecker(fsn);

    cacheManager.modifyDirective(directive, pc, flags);
    fsn.getEditLog().logModifyCacheDirectiveInfo(directive, logRetryCache);
  }

But when SBN replay the log ,it will invoke FSImageSerialization#readCacheDirectiveInfo as a absolute expiryTime.It will result in the inconsistency .

  public static CacheDirectiveInfo readCacheDirectiveInfo(DataInput in)
      throws IOException {
    CacheDirectiveInfo.Builder builder =
        new CacheDirectiveInfo.Builder();
    builder.setId(readLong(in));
    int flags = in.readInt();
    if ((flags & 0x1) != 0) {
      builder.setPath(new Path(readString(in)));
    }
    if ((flags & 0x2) != 0) {
      builder.setReplication(readShort(in));
    }
    if ((flags & 0x4) != 0) {
      builder.setPool(readString(in));
    }
    if ((flags & 0x8) != 0) {
      builder.setExpiration(
          CacheDirectiveInfo.Expiration.newAbsolute(readLong(in)));
    }
    if ((flags & ~0xF) != 0) {
      throw new IOException("unknown flags set in " +
          "ModifyCacheDirectiveInfoOp: " + flags);
    }
    return builder.build();
  }

In other words， fsn.getEditLog().logModifyCacheDirectiveInfo(directive, logRetryCache) may serial a relative expiry time,But builder.setExpiration(CacheDirectiveInfo.Expiration.newAbsolute(readLong(in))) read it as a absolute expiryTime.