Normal Description
There are a few issues with speculative retry:
1. Time unit bugs. These are from ColumnFamilyStore (v3.0.10):
The left hand side is in nanos, as the name suggests, while the right hand side is in millis.
sampleLatencyNanos = DatabaseDescriptor.getReadRpcTimeout() / 2;
Here coordinatorReadLatency is already in nanos and we shouldn't multiple the value by 1000. This was a regression in 8896a70 when we switch metrics library and the two libraries use different time units.
sampleLatencyNanos = (long) (metric.coordinatorReadLatency.getSnapshot().getValue(retryPolicy.threshold()) * 1000d);
2. Confusing overload protection and retry delay. As the name "sampleLatencyNanos" suggests, it should be used to keep the actually sampled read latency. However, we assign it the retry threshold in the case of CUSTOM. Then we compare the retry threshold with read timeout (defaults to 5000ms). This means, if we use speculative_retry=10ms for the table, we won't be able to avoid being overloaded. We should compare the actual read latency with the read timeout for overload protection. See line 450 of ColumnFamilyStore.java and line 279 of AbstractReadExecutor.java.
My proposals are:
a. We use sampled p99 delay and compare it with a customizable threshold (-Dcassandra.overload.threshold) for overload detection.
b. Introduce another variable retryDelayNanos for waiting time before retry. This is the value from table setting (PERCENTILE or CUSTOM).