Tablet throttler

VTTablet runs a cooperative throttling service. This service probes the shard's MySQL topology and observes health, measure by replication lag, or by another metric delivered by custom query, on servers. This throttler is derived from GitHub's freno.

Why throttler: maintaining shard health via low replication lag #

Vitess uses MySQL with asynchronous or semi-synchronous replication. In these modes, each shard has a primary instance that applies changes and logs them to the binary log. The replicas for that shard will get binary log entries from the primary, potentially acknowledge them (if semi-synchronous replication is enabled), and apply them. A running replica normally applies the entries as soon as possible, unless it is stopped or configured to delay. However, if the replica is busy, then it may not have the resources to apply events in a timely fashion, and can therefore start lagging. For example, if the replica is serving traffic, it may lack the necessary disk I/O or CPU to avoid lagging behind the primary.

Maintaining low replication lag is important in production for two reasons:

  • A lagging replica may not be representative of the data on the primary. Reads from the replica reflect data that is not consistent with the data on the primary. This is noticeable on web services following read-after-write from the replica, and this can produce results not reflecting the write.
  • An up-to-date replica makes for a good failover experience. If all replicas are lagging, then a failover process must choose between waiting for a replica to catch up or losing data.

Some common database operations include mass writes to the database, including the following:

  • Online schema migrations duplicating entire tables
  • Mass population of columns, such as populating the new column with derived values following an ADD COLUMN migration
  • Purging of old data
  • Purging of tables as part of safe table DROP operation

Other operations include mass reads from the database:

  • An ETL reading content of entire tables
  • VReplication scanning an entire keyspace data and binary logs

These operations can easily incur replication lag. However, these operations are typically not time-limited. It is possible to rate-limit them to reduce database load.

This is where a throttler becomes useful. A throttler can detect when replication lag is low, a cluster is healthy, and operations can proceed. It can also detect when replication lag is high and advise applications to hold the next operation.

Applications are expected to break down their tasks into small sub-tasks. For example, instead of deleting 1,000,000 rows, an application should only delete 50 at a time. Between these sub-tasks, the application should check in with the throttler.

The throttler is only intended for use with operations such as the above mass write/read cases. It should not be used for ongoing, normal OLTP queries.

Throttler overview #

Each vttablet runs an internal throttler service, and provides API endpoints to the throttler. Each tablet, including the primary, measures its own "self" health, discussed later.

Cluster health: #

In addition, the primary tablet is responsible for the overall health of the cluster/shard:

  • The throttler confirms it is still the primary tablet for its shard.
  • Every 10sec, the throttler uses the topology server to refresh the shard's tablets list.
  • The throttler probes all REPLICA tablets for their replication lag. This is done by querying the _vt.heartbeat table.
    • The throttler begins in dormant probe mode. As long as no application or client is actually looking for metrics, it probes the servers at multi-second intervals.
    • When applications check for throttle advice, the throttler begins probing servers in subsecond intervals. It reverts to dormant probe mode if no requests are made in the duration of 1min.
  • The throttler aggregates the last probed values from all relevant tablets. This is _the cluster's metric _.

The cluster's metric is only as accurate as the following metrics:

  • The probe interval
  • The heartbeat injection interval
  • The aggregation interval

The error margin equals approximately the sum of the above values, plus additional overhead. The defaults for these intervals are as follows:

  • Probe interval: 100ms
  • Aggregation interval: 100ms
  • Heartbeat interval: 250ms

The user may override the heartbeat interval by sending -heartbeat_interval flag to vttablet.

Thus, the aggregated interval can be off, by default, by some 500ms. This makes it inaccurate for evaluations that require high resolution lag evaluation. This resolution is sufficient for throttling purposes.

Self health #

Each tablet runs a local health check against its backend database, again in the form of evaluating replication lag from _vt.heartbeat. Intervals are identical to the cluster health interval illustrated above.

Response codes #

The throttler allows clients and applications to check for throttle advice. The check is an HTTP request, HEAD method, or GET method. Throttler returns one of the following HTTP response codes as an answer:

  • 200 (OK): The application may write to the data store. This is the desired response.
  • 404 (Not Found): The check contains an unknown metric name. This can take place immediately upon startup or immediately after failover, and should resolve within 10 seconds.
  • 417 (Expectation Failed): The requesting application is explicitly forbidden to write. The throttler does not implement this at this time.
  • 429 (Too Many Requests): Do not write. A normal, expected state indicating there is replication lag. This is the hint for applications or clients to withhold writes.
  • 500 (Internal Server Error): An internal error has occurred. Do not write.

Normally, apps will see either 200 or 429. An app should only ever proceed to write to the database when it receives a 200 response code.

The throttler chooses the response by comparing the replication lag with a pre-defined threshold. If the lag is lower than the threshold, response can be 200 (OK). If the lag is higher than the threshold, the response would be 429 (Too Many Requests).

The throttler only collects and evaluates lag on a set of predefined tablet types. By default, this tablet type set is REPLICA. See Configuration.

When the throttler sees no relevant replicas in the shard, it allows writes by responding with HTTP 200 OK.

Custom metrics & queries #

The default behavior is to measure replication lag and throttle based on that lag. Vitess allows the user to use custom metrics and thresholds for throttling.

Vitess only supports gauges for custom metrics: the user may define a query which returns a gauge value, an absolute metric by which Vitess can throttle. See #Configuration, below.

Configuration #

  • The throttler is currently disabled by default. Use the vttablet option -enable-lag-throttler to enable the throttler. When the throttler is disabled, it still serves /throttler/check and /throttler/check-self API endpoints, and responds with HTTP 200 OK to all requests. When the throttler is enabled, it implicitly also runs heartbeat injections.
  • Use the vttablet flag -throttle_threshold to set a lag threshold value. The default threshold is 1sec and is set upon tablet startup. For example, to set a half-second lag threshold, use the flag -throttle_threshold=0.5s.
  • To set the tablet types that the throttler queries for lag, use the vttablet flag -throttle_tablet_types="replica,rdonly". The default tablet type is replica; this type is always implicitly included in the tablet types list. You may add any other tablet type. Any type not specified is ignored by the throttler.
  • To override the default lag evaluation, and measure a different metric, use -throttle_metrics_query. The query must be either of these forms:
    • SHOW GLOBAL STATUS LIKE '<metric>'
    • SHOW GLOBAL VARIABLES LIKE '<metric>'
    • SELECT <single-column> FROM ..., expecting single column, single row result
  • To override the throttle threshold, use -throttle_metrics_threshold. Floating point values are accepted.
  • Use -throttle_check_as_check_self to implicitly reroute any /throttler/check call into /throttler/check-self. This makes sense when the user supplies a custom query, and where the user wishes to throttle writes to the cluster based on the primary tablet's health, rather than the overall health of the cluster.

An example for custom query & threshold setup, using the MySQL metrics Threads_running (number of threads actively executing a query at a given time) on the primary, might look like:

$ vttablet
  -throttle_metrics_query "show global status like 'threads_running'"
  -throttle_metrics_threshold 150
  -throttle_check_as_check_self

API & usage #

Applications use these API endpoints:

  • /throttler/check, for apps that wish to write mass amounts of data to a shard, and wish to maintain the overall health of the shard.

  • /throttler/check-self, for apps that wish to perform some operation (e.g. a massive read) on a specific tablet and only wish to maintain the health of that tablet.

  • Applications may indicate their identity via ?app=<name> parameter.

  • Applications may also declare themselves to be low priority via ?p=low param. Managed online schema migrations (gh-ost, pt-online-schema-change) do so, as does the table purge process.

Examples:

  • gh-ost uses this throttler endpoint: /throttler/check?app=gh-ost&p=low
  • A data backfill application will identify as such, and use normal priority: /throttler/check?app=my_backfill (priority not indicated in URL therefore assumed to be normal)
  • An app reading a massive amount of data directly from a replica tablet will use /throttler/check-self?app=my_data_reader

A HEAD request is sufficient. A GET request also provides a JSON output. For example:

  • {"StatusCode":200,"Value":0.207709,"Threshold":1,"Message":""}
  • {"StatusCode":429,"Value":3.494452,"Threshold":1,"Message":"Threshold exceeded"}
  • {"StatusCode":404,"Value":0,"Threshold":0,"Message":"No such metric"}

In the first two above examples we can see that the tablet is configured to throttle at 1sec

Tablet also provides /throttler/status endpoint. This is useful for monitoring and management purposes.

Example: Healthy primary tablet

The following command gets throttler status on a primary tablet hosted on tablet1, serving on port 15100.

$ curl -s http://tablet1:15100/throttler/status | jq .

This API call returns the following JSON object:

{
  "Keyspace": "commerce",
  "Shard": "80-c0",
  "IsLeader": true,
  "IsOpen": true,
  "IsDormant": false,
  "AggregatedMetrics": {
    "mysql/self": {
      "Value": 0.749837
    },
    "mysql/shard": {
      "Value": 0.749887
    }
  },
  "MetricsHealth": {
    "mysql/self": {
      "LastHealthyAt": "2021-01-24T19:03:19.141933727+02:00",
      "SecondsSinceLastHealthy": 0
    },
    "mysql/shard": {
      "LastHealthyAt": "2021-01-24T19:03:19.141974429+02:00",
      "SecondsSinceLastHealthy": 0
    }
  }
}

The primary tablet serves two types of metrics:

  • mysql/shard: an aggregated lag on relevant replicas in this shard. This is the metric to check when writing massive amounts of data to this server.
  • mysql/self: the health of the specific primary MySQL server backed by this tablet.

"IsLeader": true indicates this tablet is active, is the primary, and is running probes. "IsDormant": false, means that an application has recently issued a check, and the throttler is probing for lag at high frequency.

Example: replica tablet

The following command gets throttler status on a replica tablet hosted on tablet2, serving on port 15100.

$ curl -s http://tablet2:15100/throttler/status | jq .

This API call returns the following JSON object:

{
  "Keyspace": "commerce",
  "Shard": "80-c0",
  "IsLeader": false,
  "IsOpen": true,
  "IsDormant": false,
  "AggregatedMetrics": {
    "mysql/self": {
      "Value": 0.346409
    }
  },
  "MetricsHealth": {
    "mysql/self": {
      "LastHealthyAt": "2021-01-24T19:04:25.038290475+02:00",
      "SecondsSinceLastHealthy": 0
    }
  }
}

The replica tablet only presents mysql/self metric (measurement of its own backend MySQL's lag). It does not serve checks for the shard in general.

Resources #