Skip to content
This repository has been archived by the owner on Jan 30, 2020. It is now read-only.

Latest commit

 

History

History
53 lines (39 loc) · 2.42 KB

fleet-scaling.md

File metadata and controls

53 lines (39 loc) · 2.42 KB
Raw

fleet and scaling

As fleet currently uses etcd for cluster wide coordination, making it scale requires minimizing the load it puts on etcd. This is true for reads, writes, and watches.

Known issues

  • Currently when fleet schedules a job all fleetds are woken up (via a watch) and then do a recursive GET on the Unit file in etcd to figure out if it should schedule a job. This is a very expensive operation.

  • With a large number of units (~500+), the fleetd agent exhibits significant CPU usage from parsing all D-Bus messages when interacting with systemd.

  • With a large number of units, fleetd exhibits significant CPU usage when

    • a) parsing the JSON-encoded representations (stored in etcd)
    • b) parsing the unit file itself (using go-systemd)

  • The agent deals very poorly with inconsistent read/write latencies with etcd (what is the actual behaviour?)

Improvements

To make fleet scale further in the future we could consider rearchitecting fleet to provide per-machine schedules (so that only those fleetds are woken up that actually have work for them). This is akin to the Thundering herd problem, but in a distributed fashion. Once such a change is in we can also drop the periodic wakeups (agent TTLs) that cause fleet wide wake-ups on a regular clock.

Ultimately, fleet should move away from using etcd as an RPC mechanism. Instead, it should use etcd only for leader election and then perform direct RPCs between the engine and agent.

Implemented quick wins

  • Disallowing (some) nodes to partake in the fleet leadership election. Again this is an expensive operation. The fewer nodes that are engaged in this election, the better. Possible downside is that if there isn't a leader at all, the cluster is inoperable. However the (usually) 5 machines running etcd are also a single point of failure. See the disable_engine config flag.

  • Making some defaults exported and allow them to be overridden. For instance fleet's tokenLimit controls how many Units are listed per "page". See the token_limit config flag.

  • Removing watches from fleet: By removing the watches from fleet we stop the entire cluster from walking up whenever a new job is to be scheduled. The downside of this change is that fleet's responsiveness is lower. See the disable_watches config flag.