kiwicom / gocql

Package gocql implements a fast and robust Cassandra client for the Go programming language.

Scylla shard-aware fork of gocql/gocql

This is a fork of gocql package that we created at Scylla. It contains extensions to tokenAwareHostPolicy supported by the Scylla 2.3 and onwards. It allows driver to select a connection to a particular shard on a host based on the token. This eliminates passing data between shards and significantly reduces latency. The protocol extension spec is available here.

There are open pull requests to merge the functionality to the upstream project:

• gocql/gocql#1210
• gocql/gocql#1211.

It also provides support for shard aware ports, a faster way to connect to all shards, details available in blogpost.

Installation

This is a drop-in replacement to gocql, it reuses the github.com/gocql/gocql import path.

Add the following line to your project go.mod file.

replace github.com/gocql/gocql => github.com/scylladb/gocql latest

and run

go mod tidy

to evaluate latest to a concrete tag.

Your project now uses the Scylla driver fork, make sure you are using the TokenAwareHostPolicy to enable the shard-awareness, continue reading for details.

Configuration

In order to make shard-awareness work, token aware host selection policy has to be enabled. Please make sure that the gocql configuration has PoolConfig.HostSelectionPolicy properly set like in the example below.

When working with a Scylla cluster, PoolConfig.NumConns option has no effect - the driver opens one connection for each shard and completely ignores this option.

c := gocql.NewCluster(hosts...)

// Enable token aware host selection policy, if using multi-dc cluster set a local DC.
fallback := gocql.RoundRobinHostPolicy()
if localDC != "" {
	fallback = gocql.DCAwareRoundRobinPolicy(localDC)
}
c.PoolConfig.HostSelectionPolicy = gocql.TokenAwareHostPolicy(fallback)

// If using multi-dc cluster use the "local" consistency levels. 
if localDC != "" {
	c.Consistency = gocql.LocalQuorum
}

// When working with a Scylla cluster the driver always opens one connection per shard, so `NumConns` is ignored.
// c.NumConns = 4

Shard-aware port

This version of gocql supports a more robust method of establishing connection for each shard by using shard aware port for native transport. It greatly reduces time and the number of connections needed to establish a connection per shard in some cases - ex. when many clients connect at once, or when there are non-shard-aware clients connected to the same cluster.

If you are using a custom Dialer and if your nodes expose the shard-aware port, it is highly recommended to update it so that it uses a specific source port when connecting.

• If you are using a custom net.Dialer, you can make your dialer honor the source port by wrapping it in a gocql.ScyllaShardAwareDialer:

  oldDialer := net.Dialer{...}
  clusterConfig.Dialer := &gocql.ScyllaShardAwareDialer{oldDialer}

• If you are using a custom type implementing gocql.Dialer, you can get the source port by using the gocql.ScyllaGetSourcePort function. An example:

  func (d *myDialer) DialContext(ctx context.Context, network, addr string) (net.Conn, error) {
      sourcePort := gocql.ScyllaGetSourcePort(ctx)
      localAddr, err := net.ResolveTCPAddr(network, fmt.Sprintf(":%d", sourcePort))
      if err != nil {
          return nil, err
      }
      d := &net.Dialer{LocalAddr: localAddr}
      return d.DialContext(ctx, network, addr)
  }

  The source port might be already bound by another connection on your system. In such case, you should return an appropriate error so that the driver can retry with a different port suitable for the shard it tries to connect to.

  • If you are using net.Dialer.DialContext, this function will return an error in case the source port is unavailable, and you can just return that error from your custom Dialer.
  • Otherwise, if you detect that the source port is unavailable, you can either return gocql.ErrScyllaSourcePortAlreadyInUse or syscall.EADDRINUSE.

For this feature to work correctly, you need to make sure the following conditions are met:

• Your cluster nodes are configured to listen on the shard-aware port (native_shard_aware_transport_port option),
• Your cluster nodes are not behind a NAT which changes source ports,
• If you have a custom Dialer, it connects from the correct source port (see the guide above).

The feature is designed to gracefully fall back to the using the non-shard-aware port when it detects that some of the above conditions are not met. The driver will print a warning about misconfigured address translation if it detects it. Issues with shard-aware port not being reachable are not reported in non-debug mode, because there is no way to detect it without false positives.

If you suspect that this feature is causing you problems, you can completely disable it by setting the ClusterConfig.DisableShardAwarePort flag to false.