Poor message balancing between replicas using Redis Sentinel.

[Grumov]

The current distribution of messages between replicas is uneven. The replica following the master in the list of servers is selected more often than others. In the configuration there is a master and two replicas:

• the first replica receives 66% of messages
• the second replica receives 34%.

I propose the following fix issue: https://github.com/StackExchange/StackExchange.Redis/pull/2553

[NickCraver]

Ah we do have an issue - great, let's discuss here and hold off on PRs that change the world :)

What does your code/example look like?

[Grumov]

My example looks like: Configuration: 3 Sentinel Redis, 1 Master Redis, 2 Replicas Redis. You can also see all Redis configs in PR tests. I write data to the master and read from replicas, distributing the load between replicas. Since one of the replicas receives a large load, it is overloaded earlier than expected. I assumed according to the Round Robin strategy that the replicas would receive the same number of commands.

var connectionString = $"{TestConfig.Current.SentinelServer}:{TestConfig.Current.SentinelTwoReplicasPortA},serviceName={ServiceOptions.ServiceName},allowAdmin=true";
using var conn = await ConnectionMultiplexer.ConnectAsync(connectionString);

conn.ConfigurationChanged += (s, e) => Log($"Configuration changed: {e.EndPoint}");

var db = conn.GetDatabase();
await db.PingAsync();

// Set string value on current primary
var expected = DateTime.Now.Ticks.ToString();
Log("Tick Key: " + expected);
var key = "anyKey";
await db.KeyDeleteAsync(key, CommandFlags.FireAndForget);
await db.StringSetAsync(key, expected);

var value = await db.StringGetAsync(key);
Assert.Equal(expected, value);

Log("Waiting for first replication check...");
// force read from replica, replication has some lag
await WaitForReplicationAsync(servers[0]).ForAwait();

for (var i = 0; i < 100; i++)
{
    await db.StringGetAsync(key, CommandFlags.PreferReplica);
}

[Grumov]

If we cannot fix the balancing, then maybe you can provide a public interface so that we can make custom server selection strategies.