Package org.apache.cassandra.service.paxos

Class ContentionStrategy

  • public class ContentionStrategy
extends java.lang.Object

    A strategy for making back-off decisions for Paxos operations that fail to make progress because of other paxos operations. The strategy is defined by four factors:

    The first three represent time periods, and may be defined dynamically based on a simple calculation over:

    • pX() recent experienced latency distribution for successful operations, e.g. p50(rw) the maximum of read and write median latencies, p999(r) the 99.9th percentile of read latencies
    • attempts the number of failed attempts made by the operation so far
    • constant a user provided floating point constant

    Their calculation may take any of these forms

  • constant $constant$[mu]s
  • dynamic constant pX() * constant
  • dynamic linear pX() * constant * attempts
  • dynamic exponential pX() * constant ^ attempts

    Furthermore, the dynamic calculations can be bounded with a min/max, like so: min[mu]s <= dynamic expr <= max[mu]s e.g.

  • 10ms <= p50(rw)*0.66
  • 10ms <= p95(rw)*1.8^attempts <= 100ms
  • 5ms <= p50(rw)*0.5

    These calculations are put together to construct a range from which we draw a random number. The period we wait for X will be drawn so that min <= X < max.

    With the constraint that max must be minDelta greater than min, but no greater than its expression-defined maximum. max will be increased up until this point, after which min will be decreased until this gap is imposed.

    The waitRandomizer property specifies the manner in which a random value is drawn from the range. It is defined using one of the following specifiers:

  • uniform
  • exp($power$) or exponential($power$)
  • qexp($power$) or qexponential($power$) or quantizedexponential($power$) The uniform specifier is self-explanatory, selecting all values in the range with equal probability. The exponential specifier draws values towards the end of the range with higher probability, raising a floating point number in the range [0..1.0) to the power provided, and translating the resulting value to a uniform value in the range. The quantized exponential specifier partitions the range into attempts buckets, then applies the pure exponential approach to draw values from [0..attempts), before drawing a uniform value from the corresponding bucket

    Finally, there is also a traceAfterAttempts property that permits initiating tracing of operations that experience a certain minimum number of failed paxos rounds due to contention. A setting of 0 or 1 will initiate a trace session after the first failed ballot.