Shard entropy refers to inconsistency among shards in a shard group. This can be due to the “eventually consistent” nature of data stored in InfluxDB Enterprise clusters or due to missing or unreachable shards. The Anti-Entropy (AE) service ensures that each data node has all the shards it owns according to the metastore and that all shards in a shard group are consistent. Missing shards are automatically repaired without operator intervention while out-of-sync shards can be manually queued for repair. This topic covers how the Anti-Entropy service works and some of the basic situations where it takes effect.
The Anti-Entropy service is a component of the
influxd service available on each of your data nodes
and can be used to ensure that each data node has all of the shards that the metastore says it owns and
that those shards are in sync with other ones within the same shard group.
If any shards are missing, the Anti-Entropy service will copy existing shards from other shard owners.
If data inconsistencies are detected among shards in a shard group, you can
invoke the Anti-Entropy service and queue the
out-of-sync shards for repair.
In the repair process, the Anti-Entropy service will sync the necessary updates from other shards
within a shard group.
By default, the service performs consistency checks every 5 minutes. The interval can be modified in the
anti-entropy.check-interval configuration setting.
The Anti-Entropy service can only address missing or inconsistent shards when there is at least one copy of the shard available. In other words, as long as new and healthy nodes are introduced, a replication factor of 2 can recover from one missing or inconsistent node; a replication factor of 3 can recover from two missing or inconsistent nodes, and so on. A replication factor of 1, which is not recommended, cannot be recovered by the Anti-Entropy service.
Symptoms of entropy
The Anti-Entropy service automatically detects and fixes missing shards, but shard inconsistencies must be manually detected and queued for repair. There are symptoms of entropy that, if seen, would indicate an entropy repair is necessary.
Different results for the same query
When running queries against an InfluxDB Enterprise cluster, each query may be routed to a different data node. If entropy affects data within the queried range, the same query will return different results depending on which node it is run against.
Query attempt 1
SELECT mean("usage_idle") WHERE time > '2018-06-06T18:00:00Z' AND time < '2018-06-06T18:15:00Z' GROUP BY time(3m) FILL(0) name: cpu time mean ---- ---- 1528308000000000000 99.11867392974537 1528308180000000000 99.15410822137049 1528308360000000000 99.14927494363032 1528308540000000000 99.1980535465783 1528308720000000000 99.18584290492262
Query attempt 2
SELECT mean("usage_idle") WHERE time > '2018-06-06T18:00:00Z' AND time < '2018-06-06T18:15:00Z' GROUP BY time(3m) FILL(0) name: cpu time mean ---- ---- 1528308000000000000 99.11867392974537 1528308180000000000 0 1528308360000000000 0 1528308540000000000 0 1528308720000000000 99.18584290492262
This indicates that data is missing in the queried time range and entropy is present.
A “flapping” dashboard means data visualizations changing when data is refreshed and pulled from a node with entropy (inconsistent data). It is the visual manifestation of getting different results from the same query.
The Anti-Entropy service runs on each data node and periodically checks its shards’ statuses relative to the next data node in the ownership list. It does this by creating a “digest” or summary of data in the shards on the node.
For example, assume there are two data nodes in your cluster:
shard1 is replicated across each.
When a status check runs,
node1 will ask
shard1 was last modified.
If the reported modification time differs from the previous check, then
node2 for a new digest of
shard1, checks for differences (performs a “diff”) between the
shard1 digest for
node2 and the local
If a difference exists,
shard1 is flagged as having entropy.
If during a status check a node determines the next node is completely missing a shard,
it immediately adds the missing shard to the repair queue.
A background routine monitors the queue and begins the repair process as new shards are added to it.
Repair requests are pulled from the queue by the background process and repaired using a
copy shard operation.
Currently, shards that are present on both nodes but contain different data are not automatically queued for repair. A user must make the request via
influxd-ctl entropy repair <shard ID>. For more information, see Detecting and repairing entropy below.
node2 from the earlier example,
node2 for a digest of
node1 diffs its own local
shard1 digest and
then creates a new digest containing only the differences (the diff digest).
The diff digest is used to create a patch containing only the data
node2 is missing.
node1 sends the patch to
node2 and instructs it to apply it.
node2 finishes applying the patch, it queues a repair for
The “node-to-node” shard repair continues until it runs on every data node that owns the shard in need of repair.
Repairs between shard owners happen in a deterministic order. This doesn’t mean repairs always start on node 1 and then follow a specific node order. Repairs are viewed at the shard level. Each shard has a list of owners and the repairs for a particular shard will happen in a deterministic order among its owners.
When the Anti-Entropy service on any data node receives a repair request for a shard, it determines which owner node is the first in the deterministic order and forwards the request to that node. The request is now queued on the first owner.
The first owner’s repair processor pulls it from the queue, detects the differences between the local copy of the shard with the copy of the same shard on the next owner in the deterministic order, then generates a patch from that difference. The first owner then makes an RPC call to the next owner instructing it to apply the patch to its copy of the shard.
Once the next owner has successfully applied the patch, it adds that shard to the Anti-Entropy repair queue. A list of “visited” nodes follows the repair through the list of owners. Each owner will check the list to detect when the repair has cycled through all owners, at which point the repair is finished.
The Anti-Entropy service does its best to avoid hot shards (shards that are currently receiving writes) because they change quickly. While write replication between shard owner nodes (with a replication factor greater than 1) typically happens in milliseconds, this slight difference is still enough to cause the appearance of entropy where there is none.
Because the Anti-Entropy service repairs only cold shards, unexpected effects can occur. Consider the following scenario:
- A shard goes cold.
- Anti-Entropy detects entropy.
- Entropy is reported by the Anti-Entropy
/statusAPI or with the
influxd-ctl entropy showcommand.
- Shard takes a write, gets compacted, or something else causes it to go hot. These actions are out of Anti-Entropy’s control.
- A repair is requested, but is ignored because the shard is now hot.
In this example, you would have to periodically request a repair of the shard until it either shows as being in the queue, being repaired, or no longer in the list of shards with entropy.
The configuration settings for the Anti-Entropy service are described in Anti-Entropy settings section of the data node configuration.
To enable the Anti-Entropy service, change the default value of the
[anti-entropy].enabled = false setting to
true in the
influxdb.conf file of each of your data nodes.
Command line tools for managing entropy
Note: The Anti-Entropy service is disabled by default and must be enabled before using these commands.
influxd-ctl entropy command enables you to manage entropy among shards in a cluster.
It includes the following subcommands:
Lists shards that are in an inconsistent state and in need of repair as well as shards currently in the repair queue.
influxd-ctl entropy show
Queues a shard for repair.
It requires a Shard ID which is provided in the
influxd-ctl entropy repair <shardID>
Repairing entropy in a shard is an asynchronous operation. This command will return quickly as it only adds a shard to the repair queue. Queuing shards for repair is idempotent. There is no harm in making multiple requests to repair the same shard even if it is already queued, currently being repaired, or not in need of repair.
Removes a shard from the repair queue.
It requires a Shard ID which is provided in the
influxd-ctl entropy kill-repair <shardID>
This only applies to shards in the repair queue. It does not cancel repairs on nodes that are in the process of being repaired. Once a repair has started, requests to cancel it are ignored.
Stopping a entropy repair for a missing shard operation is not currently supported. It may be possible to stop repairs for missing shards with the
InfluxDB Anti-Entropy API
The Anti-Entropy service uses an API for managing and monitoring entropy. Details on the available API endpoints can be found in The InfluxDB Anti-Entropy API.
Common use cases for the Anti-Entropy service include detecting and repairing entropy, replacing unresponsive data nodes, replacing data nodes for upgrades and maintenance, and eliminating entropy in active shards.
Detecting and repairing entropy
Periodically, you may want to see if shards in your cluster have entropy or are
inconsistent with other shards in the shard group.
influxd-ctl entropy show command to list all shards with detected entropy:
influxd-ctl entropy show Entropy ========== ID Database Retention Policy Start End Expires Status 21179 statsdb 1hour 2017-10-09 00:00:00 +0000 UTC 2017-10-16 00:00:00 +0000 UTC 2018-10-22 00:00:00 +0000 UTC diff 25165 statsdb 1hour 2017-11-20 00:00:00 +0000 UTC 2017-11-27 00:00:00 +0000 UTC 2018-12-03 00:00:00 +0000 UTC diff
Then use the
influxd-ctl entropy repair command to add the shards with entropy
to the repair queue:
influxd-ctl entropy repair 21179 Repair Shard 21179 queued influxd-ctl entropy repair 25165 Repair Shard 25165 queued
Check on the status of the repair queue with the
influxd-ctl entropy show command:
influxd-ctl entropy show Entropy ========== ID Database Retention Policy Start End Expires Status 21179 statsdb 1hour 2017-10-09 00:00:00 +0000 UTC 2017-10-16 00:00:00 +0000 UTC 2018-10-22 00:00:00 +0000 UTC diff 25165 statsdb 1hour 2017-11-20 00:00:00 +0000 UTC 2017-11-27 00:00:00 +0000 UTC 2018-12-03 00:00:00 +0000 UTC diff Queued Shards: [21179 25165]
Replacing an unresponsive data node
If a data node suddenly disappears due to a catastrophic hardware failure or for any other reason, as soon as a new data node is online, the Anti-Entropy service will copy the correct shards to the new replacement node. The time it takes for the copying to complete is determined by the number of shards to be copied and how much data is stored in each.
_View the Replacing Data Nodes documentation for instructions on replacing data nodes in your InfluxDB Enterprise cluster._
Replacing a machine that is running a data node
Perhaps you are replacing a machine that is being decommissioned, upgrading hardware, or something else entirely. The Anti-Entropy service will automatically copy shards to the new machines.
Once you have successfully run the
influxd-ctl update-data command, you are free
to shut down the retired node without causing any interruption to the cluster.
The Anti-Entropy process will continue copying the appropriate shards from the
remaining replicas in the cluster.
Fixing entropy in active shards
In rare cases, the currently active shard, or the shard to which new data is
currently being written, may find itself with inconsistent data.
Because the Anti-Entropy process can’t write to hot shards, you must stop writes to the new
shard using the
influxd-ctl truncate-shards command,
then add the inconsistent shard to the entropy repair queue:
# Truncate hot shards influxd-ctl truncate-shards # Show shards with entropy influxd-ctl entropy show Entropy ========== ID Database Retention Policy Start End Expires Status 21179 statsdb 1hour 2018-06-06 12:00:00 +0000 UTC 2018-06-06 23:44:12 +0000 UTC 2018-12-06 00:00:00 +0000 UTC diff # Add the inconsistent shard to the repair queue influxd-ctl entropy repair 21179
Queued repairs are not being processed
The primary reason a repair in the repair queue isn’t being processed is because it went “hot” after the repair was queued. The Anti-Entropy service only repairs cold shards or shards that are not currently being written to. If the shard is hot, the Anti-Entropy service will wait until it goes cold again before performing the repair.
If the shard is “old” and writes to it are part of a backfill process, you simply
have to wait the until the backfill process is finished. If the shard is the active
shard, you can
truncate-shards to stop writes to active shards. This process is
Anti-Entropy log messages
Below are common messages output by Anti-Entropy along with what they mean.
Indicates that the Anti-Entropy process has begun the status check process.
Indicates that the Anti-Entropy process has skipped a status check on shards because they are currently hot.