Managing Active Anti-Entropy

Riak’s active anti-entropy (AAE) subsystem is a set of background processes that repair object inconsistencies stemming from missing or divergent object values across nodes. Riak operators can turn AAE on and off and configure and monitor its functioning.

Enabling Active Anti-Entropy

Whether AAE is currently enabled in a node is determined by the value of the anti_entropy parameter in the node’s configuration files.

In Riak versions 2.0 and later, AAE is turned on by default.

anti_entropy = active
{riak_kv, [

  {anti_entropy, {on, []}},

    %% More riak_kv settings...
]}

For monitoring purposes, you can also activate AAE debugging, which provides verbose debugging message output:

anti_entropy = active-debug
{riak_kv, [

    %% With debugging
    {anti_entropy, {on, [debug]}},

    %% More riak_kv settings...
]}

Remember that you will need to restart the node for any configuration-related changes to take effect.

Disabling Active Anti-Entropy

Alternatively, AAE can be switched off if you would like to repair object inconsistencies using read repair alone:

anti_entropy = passive
{riak_kv, [

    %% AAE turned off
    {anti_entropy, {off, []}},

    %% More riak_kv settings...
]}

If you would like to reclaim the disk space used by AAE operations, you must manually delete the directory in which AAE-related data is stored in each node.

rm -Rf <path_to_riak_node>/data/anti_entropy/*

The default directory for AAE data is ./data/anti_entropy, as in the example above, but this can be changed. See the section below titled Data Directory.

Remember that you will need to restart the node for any configuration-related changes to take effect.

The directory deletion method above can also be used to force a rebuilding of hash trees.

Monitoring AAE

Riak’s command-line interface includes a command that provides insight into AAE-related processes and performance:

riak-admin aae-status

When you run this command in a node, the output will look like this (shortened for the sake of brevity):

================================== Exchanges ==================================
Index                                              Last (ago)    All (ago)
-------------------------------------------------------------------------------
0                                                  19.0 min      20.3 min
22835963083295358096932575511191922182123945984    18.0 min      20.3 min
45671926166590716193865151022383844364247891968    17.3 min      19.8 min
68507889249886074290797726533575766546371837952    16.5 min      18.3 min
91343852333181432387730302044767688728495783936    15.8 min      17.3 min
...

================================ Entropy Trees ================================
Index                                              Built (ago)
-------------------------------------------------------------------------------
0                                                  5.7 d
22835963083295358096932575511191922182123945984    5.6 d
45671926166590716193865151022383844364247891968    5.5 d
68507889249886074290797726533575766546371837952    4.3 d
91343852333181432387730302044767688728495783936    4.8 d

================================ Keys Repaired ================================
Index                                                Last      Mean      Max
-------------------------------------------------------------------------------
0                                                     0         0         0
22835963083295358096932575511191922182123945984       0         0         0
45671926166590716193865151022383844364247891968       0         0         0
68507889249886074290797726533575766546371837952       0         0         0
91343852333181432387730302044767688728495783936       0         0         0

Each of these three tables contains information for each vnode in your cluster in these three categories:

Category Measures Description
Exchanges Last When the most recent exchange between a data partition and one of its replicas was performed
All How long it has been since a partition exchanged with all of its replicas
Entropy Trees Built When the hash trees for a given partition were created
Keys Repaired Last The number of keys repaired during all key exchanges since the last node restart
Mean The mean number of keys repaired during all key exchanges since the last node restart
Max The maximum number of keys repaired during all key exchanges since the last node restart

All AAE status information obtainable using the riak-admin aae-status command is stored in-memory and is reset when a node is restarted with the exception of hash tree build information, which is persisted on disk (because hash trees themselves are persisted on disk).

Configuring AAE

Riak’s configuration files enable you not just to turn AAE on and off but also to fine-tune your cluster’s use of AAE, e.g. how much memory AAE processes should consume, how frequently specific processes should be run, etc.

Data Directory

By default, data related to AAE operations is stored in the ./data/anti_entropy directory in each Riak node. This can be changed by setting the anti_entropy.data_dir parameter to a different value.

Throttling

AAE has a built-in throttling mechanism that can insert delays between AAE repair operations when vnode mailboxes reach the length specified by the search.anti_entropy.throttle.$tier.delay parameter (more on that in the section below). Throttling can be switched on and off using the search.anti_entropy.throttle parameter. The default is on.

Throttling Tiers

If you activate AAE throttling, you can use tiered throttling to establish a series of vnode mailbox-size thresholds past which a user-specified time delay should be observed. This enables you to establish, for example, that a delay of 10 milliseconds should be observed if the mailbox of any vnode reaches 50 messages.

The general form for setting tiered throttling is as follows:

search.anti_entropy.throttle.$tier.delay
search.anti_entropy.throttle.$tier.solrq_queue_length

In the above example, $tier should be replaced with the desired name for that tier (e.g. tier1, large_mailbox_tier, etc). If you choose to set throttling tiers, you will need to set the mailbox size for one of the tiers to 0. Both the .solrq_queue_length and .delay parameters must be set for each tier.

Below is an example configuration for three tiers, with mailbox sizes of 0, 50, and 100 and time delays of 5, 10, and 15 milliseconds, respectively:

search.anti_entropy.throttle.tier1.solrq_queue_length = 0
search.anti_entropy.throttle.tier1.delay = 5ms
search.anti_entropy.throttle.tier2.solrq_queue_length = 50
search.anti_entropy.throttle.tier2.delay = 10ms
search.anti_entropy.throttle.tier3.solrq_queue_length = 100
search.anti_entropy.throttle.tier3.delay = 15ms

Bloom Filters

Bloom filters are mechanisms used to prevent reads that are destined to fail because no object exists in the location that they’re querying. Using bloom filters can improve reaction time for some queries, but entail a small general performance cost. You can switch bloom filters on and off using the anti_entropy.bloomfilter parameter.

Trigger Interval

The anti_entropy.trigger_interval setting determines how often Riak’s AAE subsystem looks for work to do, e.g. building or expiring hash trees, triggering information exchanges between nodes, etc. The default is every 15 seconds (15s). Raising this value may save resources, but at a slightly higher risk of data corruption.

Hash Trees

As a fallback measure in addition to the normal operation of AAE on-disk hash trees, Riak periodically clears and regenerates all hash trees stored on disk to ensure that hash trees correspond to the key/value data stored in Riak. This enables Riak to detect silent data corruption resulting from disk failure or faulty hardware. The anti_entropy.tree.expiry setting enables you to determine how often that takes place. The default is once a week (1w). You can set up this process to run once a day (1d), twice a day (12h), once a month (4w), and so on.

In addition to specifying how often Riak expires hash trees after they are built, you can also specify how quickly and how many hash trees are built. You can set the frequency using the anti_entropy.tree.build_limit.per_timespan parameter, for which the default is every hour (1h); the number of hash tree builds is specified by anti_entropy.tree.build_limit.number, for which the default is 1.

Write Buffer Size

While you are free to choose the backend for data storage in Riak, background AAE processes use LevelDB. You can adjust the size of the write buffer used by LevelDB for hash tree generation using the anti_entropy.write_buffer_size parameter. The default is 4MB.

Open Files and Concurrency Limits

The anti_entropy.concurrency_limit parameter determines how many AAE cross-node information exchanges or hash tree builds can happen concurrently. The default is 2.

The anti_entropy.max_open_files parameter sets an open-files limit for AAE-related background tasks, analogous to open files limit settings used in operating systems. The default is 20.

Riak’s AAE subsystem works to repair object inconsistencies both with for normal key/value objects as well as data related to Riak Search. In particular, AAE acts on indexes stored in Solr, the search platform that drives Riak Search. Implementation details for AAE and Search can be found in the Search Details documentation.

You can check on the status of Search-related AAE using the following command:

riak-admin search aae-status

The output from that command can be interpreted just like the output discussed in the section on monitoring above.