Slow start modeΒΆ

Slow start mode is a configuration setting in Envoy to progressively increase amount of traffic for newly added upstream endpoints. With no slow start enabled Envoy would send a proportional amount of traffic to new upstream endpoints. This could be undesirable for services that require warm up time to serve full production load and could result in request timeouts, loss of data and deteriorated user experience.

Slow start mode is a mechanism that affects load balancing weight of upstream endpoints and can be configured per upstream cluster. Currently, slow start is supported in Round Robin and Least Request load balancer types.

Users can specify a slow start window parameter (in seconds), which is the duration of slow start mode per endpoint. During slow start window, load balancing weight of a particular endpoint will be scaled with time factor, e.g.:

\[NewWeight = {Weight}*{max(MinWeightPercent,{TimeFactor}^\frac{1}{Aggression})}\]

where,

\[TimeFactor = \frac{max(TimeSinceStartInSeconds,1)}{SlowStartWindowInSeconds}\]

As time progresses, more and more traffic would be sent to endpoint within slow start window.

MinWeightPercent parameter specifies the minimum percent of origin weight to make sure the EDF scheduler has a reasonable deadline, default is 10%.

Aggression parameter non-linearly affects endpoint weight and represents the speed of ramp-up. By tuning aggression parameter, one could achieve polynomial or exponential speed for traffic increase. Below simulation demonstrates how various values for aggression affect traffic ramp-up:

../../../../_images/slow_start_aggression.svg

Whenever a slow start window duration elapses, upstream endpoint exits slow start mode and gets regular amount of traffic according to load balancing algorithm. Its load balancing weight will no longer be scaled with runtime bias and aggression. Endpoint could also exit slow start mode in case it leaves the cluster. Endpoint could further re-enter slow start mode, in case it has transitioned from unhealthy to healthy state via an active healchecking.

To reiterate, endpoint enters slow start mode:
  • If no active healthcheck is configured per cluster, immediately upon joining the cluster.

  • In case an active healthcheck is configured per cluster, when the endpoint has transitioned from unhealthy to healthy state via an active healcheck.

Endpoint exits slow start mode when:
  • It leaves the cluster.

  • It does not pass an active healthcheck configured per cluster. Endpoint could re-enter slow start, if it passes an active healthcheck.

It is not recommended enabling slow start mode in low traffic or high number of endpoints scenarios, potential drawbacks would be:
  • Endpoint starvation, where endpoint has low probability to receive a request either due to low traffic or high number of total endpoints.

  • Spurious (non-gradual) increase of traffic per endpoint, whenever a starving endpoint receives a request and sufficient time has passed within slow start window, its load balancing weight will increase non linearly due to time factor.

Below is an example of how result load balancing weight would look like for endpoints in same priority with Round Robin Loadbalancer type, slow start window of 60 seconds, no active healthcheck and 1.0 aggression. Once endpoints E1 and E2 exit slow start mode, their load balancing weight remains constant:

../../../../_images/slow_start_example.svg