Anti-entropy service in InfluxDB Enterprise

Introduction

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 guide covers how AE works and some of the basic situations where it takes effect.

Concepts

The anti-entropy service is part of the influxd process running on each data node that ensures the node has all the shards the metastore says it owns and that those shards are in sync with others in the same shard group. If any shards are missing, the AE service will copy existing shards from other shard owners. If data inconsistencies are detected among shards in a shard group, you can envoke the AE process and queue the out-of-sync shards for repair. In the repair process, AE will sync the necessary updates from other shards in the same shard group.

By default, the service checks every 5 minutes, as configured in the anti-entropy.check-interval 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 cannot be recovered by the anti-entropy service.

Symptoms of entropy

The AE process 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.

Flapping dashboards

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.

Flapping dashboard

Technical details

Detecting entropy

The AE 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.

As an example, assume there are two data nodes in your cluster: node1 and node2. Both node1 and node2 own shard1 so shard1 is replicated across each.

When a status check runs, node1 will ask node2 when shard1 was last modified. If the reported modification time is different than it was in the previous check, node1 will ask node2 for a new digest of shard1. node1 then checks for differences (performs a “diff”) between node2’s shard1 digest and its own local digest for shard1. If there’s a difference, shard1 is flagged as having entropy.

Repairing 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 request 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>. More info below

Using node1 and node2 from the example abovenode1 asks node2 for a digest of shard1. node1 diffs its own local shard1 digest and node2’s shard1 digest, 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. Once node2 finishes applying the patch, it queues a repair for shard1 locally.

The “node-to-node” shard repair continues until it runs on every data node that owns the shard in need of repair.

Repair order

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 AE 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 its AE 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.

Hot shards

The AE 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. AE is designed and intended to repair cold shards.

This can sometimes have some unexpected effects. For example:

  • A shard goes cold.
  • AE detects entropy.
  • Entropy reported by the AE /status API or with the influxd-ctl entropy show command.
  • Shard takes a write, gets compacted, or something else causes it to go hot. These actions are out of AE’s control.
  • A repair is requested, but ignored because the shard is now hot.

In the scenario above, you will 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.

Configuration

Anti-entropy configuration options are available in [anti-entropy] section of your influxdb.conf.

Command line tools for managing entropy

The influxd-ctl entropy command enables you to manage entropy among shards in a cluster. It includes the following subcommands:

show

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

repair

Queues a shard for repair. It requires a Shard ID which is provided in the show output.

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.

kill-repair

Removes a shard from the repair queue. It requires a Shard ID which is provided in the show output.

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 influxd-ctl kill-copy-shard command.

Scenarios

This section covers some of the common use cases for the anti-entropy service.

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. Use the 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 AE 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

Troubleshooting

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. AE can only repair cold shards, or shards that are not currently being written to. If the shard is hot, AE 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 outlined above.

AE log messages

Below are common messages output by AE along with what they mean.

Checking status

Indicates that the AE process has begun the status check process.

Skipped shards

Indicates that the AE process has skipped a status check on shards because they are currently hot.

Changes to the AE Service in v1.6

  • New entropy command in the influxd-ctl cluster management utility that includes show, repair, and kill-repair subcommands.
  • New /repair API (Documentation Coming).
  • New /cancel-repair API (Documentation Coming).
  • Updated /status API that now includes a list of shards waiting in the repair queue and a list of shards currently being repaired.
  • New repair order.
  • Repairs are now “push” instead of “pull”. In v1.5, repairs of missing shards were done with a “pull” of the shard from another node. The AE service would notice a shard missing and choose another owner to copy from. In 1.6, it happens in the deterministic order described above.

This documentation is open source. See a typo? Please, open an issue.


Need help getting up and running? Get Support