RGW Dynamic Bucket Index Resharding

New in version Luminous.

A large bucket index can lead to performance problems. In order to address this problem we introduced bucket index sharding. Until Luminous, changing the number of bucket shards (resharding) needed to be done offline. Starting with Luminous we support online bucket resharding.

Each bucket index shard can handle its entries efficiently up until reaching a certain threshold number of entries. If this threshold is exceeded the system can suffer from performance issues. The dynamic resharding feature detects this situation and automatically increases the number of shards used by the bucket index, resulting in a reduction of the number of entries in each bucket index shard. This process is transparent to the user. Write I/Os to the target bucket are blocked and read I/Os are not during resharding process.

By default dynamic bucket index resharding can only increase the number of bucket index shards to 1999, although this upper-bound is a configuration parameter (see Configuration below). When possible, the process chooses a prime number of bucket index shards to spread the number of bucket index entries across the bucket index shards more evenly.

The detection process runs in a background process that periodically scans all the buckets. A bucket that requires resharding is added to the resharding queue and will be scheduled to be resharded later. The reshard thread runs in the background and execute the scheduled resharding tasks, one at a time.

Multisite

Prior to the Reef release, RGW does not support dynamic resharding in a multisite environment. For information on dynamic resharding, see Resharding in the RGW multisite documentation.

Configuration

Enable/Disable dynamic bucket index resharding:

rgw_dynamic_resharding: true/false, default: true

Configuration options that control the resharding process:

rgw_max_objs_per_shard: maximum number of objects per bucket index shard before resharding is triggered, default: 100000 objects
rgw_max_dynamic_shards: maximum number of shards that dynamic bucket index resharding can increase to, default: 1999
rgw_reshard_bucket_lock_duration: duration, in seconds, of lock on bucket obj during resharding, default: 360 seconds (i.e., 6 minutes)
rgw_reshard_thread_interval: maximum time, in seconds, between rounds of resharding queue processing, default: 600 seconds (i.e., 10 minutes)
rgw_reshard_num_logs: number of shards for the resharding queue, default: 16

Admin commands

Add a bucket to the resharding queue

# radosgw-admin reshard add --bucket <bucket_name> --num-shards <new number of shards>

List resharding queue

# radosgw-admin reshard list

Process tasks on the resharding queue

# radosgw-admin reshard process

Bucket resharding status

# radosgw-admin reshard status --bucket <bucket_name>

The output is a json array of 3 objects (reshard_status, new_bucket_instance_id, num_shards) per shard.

For example, the output at different Dynamic Resharding stages is shown below:

1. Before resharding occurred:

[
  {
      "reshard_status": "not-resharding",
      "new_bucket_instance_id": "",
      "num_shards": -1
  }
]

2. During resharding:

[
  {
      "reshard_status": "in-progress",
      "new_bucket_instance_id": "1179f470-2ebf-4630-8ec3-c9922da887fd.8652.1",
      "num_shards": 2
  },
  {
      "reshard_status": "in-progress",
      "new_bucket_instance_id": "1179f470-2ebf-4630-8ec3-c9922da887fd.8652.1",
      "num_shards": 2
  }
]

3, After resharding completed:

[
  {
      "reshard_status": "not-resharding",
      "new_bucket_instance_id": "",
      "num_shards": -1
  },
  {
      "reshard_status": "not-resharding",
      "new_bucket_instance_id": "",
      "num_shards": -1
  }
]

Cancel pending bucket resharding

Note: Ongoing bucket resharding operations cannot be cancelled.

# radosgw-admin reshard cancel --bucket <bucket_name>

Manual immediate bucket resharding

# radosgw-admin bucket reshard --bucket <bucket_name> --num-shards <new number of shards>

When choosing a number of shards, the administrator should keep a number of items in mind. Ideally the administrator is aiming for no more than 100000 entries per shard, now and through some future point in time.

Additionally, bucket index shards that are prime numbers tend to work better in evenly distributing bucket index entries across the shards. For example, 7001 bucket index shards is better than 7000 since the former is prime. A variety of web sites have lists of prime numbers; search for “list of prime numbers” withy your favorite web search engine to locate some web sites.

Troubleshooting

Clusters prior to Luminous 12.2.11 and Mimic 13.2.5 left behind stale bucket instance entries, which were not automatically cleaned up. The issue also affected LifeCycle policies, which were not applied to resharded buckets anymore. Both of these issues can be worked around using a couple of radosgw-admin commands.

Stale instance management

List the stale instances in a cluster that are ready to be cleaned up.

# radosgw-admin reshard stale-instances list

Clean up the stale instances in a cluster. Note: cleanup of these instances should only be done on a single site cluster.

# radosgw-admin reshard stale-instances rm

Lifecycle fixes

For clusters that had resharded instances, it is highly likely that the old lifecycle processes would have flagged and deleted lifecycle processing as the bucket instance changed during a reshard. While this is fixed for newer clusters (from Mimic 13.2.6 and Luminous 12.2.12), older buckets that had lifecycle policies and that have undergone resharding will have to be manually fixed.

The command to do so is:

# radosgw-admin lc reshard fix --bucket {bucketname}

As a convenience wrapper, if the --bucket argument is dropped then this command will try and fix lifecycle policies for all the buckets in the cluster.

Object Expirer fixes

Objects subject to Swift object expiration on older clusters may have been dropped from the log pool and never deleted after the bucket was resharded. This would happen if their expiration time was before the cluster was upgraded, but if their expiration was after the upgrade the objects would be correctly handled. To manage these expire-stale objects, radosgw-admin provides two subcommands.

Listing:

# radosgw-admin objects expire-stale list --bucket {bucketname}

Displays a list of object names and expiration times in JSON format.

Deleting:

# radosgw-admin objects expire-stale rm --bucket {bucketname}

Initiates deletion of such objects, displaying a list of object names, expiration times, and deletion status in JSON format.