Supported load balancers

When a filter needs to acquire a connection to a host in an upstream cluster, the cluster manager uses a load balancing policy to determine which host is selected. The load balancing policies are pluggable and are specified on a per upstream cluster basis in the configuration. Note that if no active health checking policy is configured for a cluster, all upstream cluster members are considered healthy, unless otherwise specified through health_status.

Weighted round robin

This is a simple policy in which each available upstream host is selected in round robin order. If weights are assigned to endpoints in a locality, then a weighted round robin schedule is used, where higher weighted endpoints will appear more often in the rotation to achieve the effective weighting.

Weighted least request

The least request load balancer uses different algorithms depending on whether hosts have the same or different weights.

all weights equal: An O(1) algorithm which selects N random available hosts as specified in the configuration (2 by default) and picks the host which has the fewest active requests (Mitzenmacher et al. has shown that this approach is nearly as good as an O(N) full scan). This is also known as P2C (power of two choices). The P2C load balancer has the property that a host with the highest number of active requests in the cluster will never receive new requests. It will be allowed to drain until it is less than or equal to all of the other hosts.
all weights not equal: If two or more hosts in the cluster have different load balancing weights, the load balancer shifts into a mode where it uses a weighted round robin schedule in which weights are dynamically adjusted based on the host’s request load at the time of selection.

In this case the weights are calculated at the time a host is picked using the following formula:

weight = load_balancing_weight / (active_requests + 1)^active_request_bias.

active_request_bias can be configured via runtime and defaults to 1.0. It must be greater than or equal to 0.0.

The larger the active request bias is, the more aggressively active requests will lower the effective weight.

If active_request_bias is set to 0.0, the least request load balancer behaves like the round robin load balancer and ignores the active request count at the time of picking.

For example, if active_request_bias is 1.0, a host with weight 2 and an active request count of 4 will have an effective weight of 2 / (4 + 1)^1 = 0.4. This algorithm provides good balance at steady state but may not adapt to load imbalance as quickly. Additionally, unlike P2C, a host will never truly drain, though it will receive fewer requests over time.

Ring hash

The ring/modulo hash load balancer implements consistent hashing to upstream hosts. Each host is mapped onto a circle (the “ring”) by hashing its address; each request is then routed to a host by hashing some property of the request, and finding the nearest corresponding host clockwise around the ring. This technique is also commonly known as “Ketama” hashing, and like all hash-based load balancers, it is only effective when protocol routing is used that specifies a value to hash on. If you want something other than the host’s address to be used as the hash key (e.g. the semantic name of your host in a Kubernetes StatefulSet), then you can specify it in the "envoy.lb" LbEndpoint.Metadata e.g.:

  filter_metadata:
    envoy.lb:
      hash_key: "YOUR HASH KEY"

This will override use_hostname_for_hashing.

Each host is hashed and placed on the ring some number of times proportional to its weight. For example, if host A has a weight of 1 and host B has a weight of 2, then there might be three entries on the ring: one for host A and two for host B. This doesn’t actually provide the desired 2:1 partitioning of the circle, however, since the computed hashes could be coincidentally very close to one another; so it is necessary to multiply the number of hashes per host—for example inserting 100 entries on the ring for host A and 200 entries for host B—to better approximate the desired distribution. Best practice is to explicitly set minimum_ring_size and maximum_ring_size, and monitor the min_hashes_per_host and max_hashes_per_host gauges to ensure good distribution. With the ring partitioned appropriately, the addition or removal of one host from a set of N hosts will affect only 1/N requests.

When priority based load balancing is in use, the priority level is also chosen by hash, so the endpoint selected will still be consistent when the set of backends is stable.

Maglev

The Maglev load balancer implements consistent hashing to upstream hosts. It uses the algorithm described in section 3.4 of this paper with a fixed table size of 65537 (see section 5.3 of the same paper). Maglev can be used as a drop in replacement for the ring hash load balancer any place in which consistent hashing is desired. Like the ring hash load balancer, a consistent hashing load balancer is only effective when protocol routing is used that specifies a value to hash on. If you want something other than the host’s address to be used as the hash key (e.g. the semantic name of your host in a Kubernetes StatefulSet), then you can specify it in the "envoy.lb" LbEndpoint.Metadata e.g.:

  filter_metadata:
    envoy.lb:
      hash_key: "YOUR HASH KEY"

This will override use_hostname_for_hashing.

The table construction algorithm places each host in the table some number of times proportional to its weight, until the table is completely filled. For example, if host A has a weight of 1 and host B has a weight of 2, then host A will have 21,846 entries and host B will have 43,691 entries (totaling 65,537 entries). The algorithm attempts to place each host in the table at least once, regardless of the configured host and locality weights, so in some extreme cases the actual proportions may differ from the configured weights. For example, if the total number of hosts is larger than the fixed table size, then some hosts will get 1 entry each and the rest will get 0, regardless of weight. Best practice is to monitor the min_entries_per_host and max_entries_per_host gauges to ensure no hosts are underrepresented or missing.

In general, when compared to the ring hash (“ketama”) algorithm, Maglev has substantially faster table lookup build times as well as host selection times (approximately 10x and 5x respectively when using a large ring size of 256K entries). While Maglev aims for minimal disruption, it is not as stable as ring hash when upstream hosts change. More keys will move position when hosts are removed (simulations show approximately double the keys will move). The amount of disruption can be minimized by increasing the table_size. With that said, for many applications including Redis, Maglev is very likely a superior drop in replacement for ring hash. The advanced reader can use this benchmark to compare ring hash versus Maglev with different parameters.

Random

The random load balancer selects a random available host. The random load balancer generally performs better than round robin if no health checking policy is configured. Random selection avoids bias towards the host in the set that comes after a failed host.