123 lines
5.8 KiB
Nix
123 lines
5.8 KiB
Nix
# This performs a full 'end-to-end' test of a multi-node CockroachDB cluster
|
|
# using the built-in 'cockroach workload' command, to simulate a semi-realistic
|
|
# test load. It generally takes anywhere from 3-5 minutes to run and 1-2GB of
|
|
# RAM (though each of 3 workers gets 2GB allocated)
|
|
#
|
|
# CockroachDB requires synchronized system clocks within a small error window
|
|
# (~500ms by default) on each node in order to maintain a multi-node cluster.
|
|
# Cluster joins that are outside this window will fail, and nodes that skew
|
|
# outside the window after joining will promptly get kicked out.
|
|
#
|
|
# To accomodate this, we use QEMU/virtio infrastructure and load the 'ptp_kvm'
|
|
# driver inside a guest. This driver allows the host machine to pass its clock
|
|
# through to the guest as a hardware clock that appears as a Precision Time
|
|
# Protocol (PTP) Clock device, generally /dev/ptp0. PTP devices can be measured
|
|
# and used as hardware reference clocks (similar to an on-board GPS clock) by
|
|
# NTP software. In our case, we use Chrony to synchronize to the reference
|
|
# clock.
|
|
#
|
|
# This test is currently NOT enabled as a continuously-checked NixOS test.
|
|
# Ideally, this test would be run by Hydra and Borg on all relevant changes,
|
|
# except:
|
|
#
|
|
# - Not every build machine is compatible with the ptp_kvm driver.
|
|
# Virtualized EC2 instances, for example, do not support loading the ptp_kvm
|
|
# driver into guests. However, bare metal builders (e.g. Packet) do seem to
|
|
# work just fine. In practice, this means x86_64-linux builds would fail
|
|
# randomly, depending on which build machine got the job. (This is probably
|
|
# worth some investigation; I imagine it's based on ptp_kvm's usage of paravirt
|
|
# support which may not be available in 'nested' environments.)
|
|
#
|
|
# - ptp_kvm is not supported on aarch64, otherwise it seems likely Cockroach
|
|
# could be tested there, as well. This seems to be due to the usage of
|
|
# the TSC in ptp_kvm, which isn't supported (easily) on AArch64. (And:
|
|
# testing stuff, not just making sure it builds, is important to ensure
|
|
# aarch64 support remains viable.)
|
|
#
|
|
# For future developers who are reading this message, are daring and would want
|
|
# to fix this, some options are:
|
|
#
|
|
# - Just test a single node cluster instead (boring and less thorough).
|
|
# - Move all CI to bare metal packet builders, and we can at least do x86_64-linux.
|
|
# - Get virtualized clocking working in aarch64, somehow.
|
|
# - Add a 4th node that acts as an NTP service and uses no PTP clocks for
|
|
# references, at the client level. This bloats the node and memory
|
|
# requirements, but would probably allow both aarch64/x86_64 to work.
|
|
#
|
|
|
|
let
|
|
|
|
# Creates a node. If 'joinNode' parameter, a string containing an IP address,
|
|
# is non-null, then the CockroachDB server will attempt to join/connect to
|
|
# the cluster node specified at that address.
|
|
makeNode = locality: myAddr: joinNode:
|
|
{ nodes, pkgs, lib, config, ... }:
|
|
|
|
{
|
|
# Bank/TPC-C benchmarks take some memory to complete
|
|
virtualisation.memorySize = 2048;
|
|
|
|
# Install the KVM PTP "Virtualized Clock" driver. This allows a /dev/ptp0
|
|
# device to appear as a reference clock, synchronized to the host clock.
|
|
# Because CockroachDB *requires* a time-synchronization mechanism for
|
|
# the system time in a cluster scenario, this is necessary to work.
|
|
boot.kernelModules = [ "ptp_kvm" ];
|
|
|
|
# Enable and configure Chrony, using the given virtualized clock passed
|
|
# through by KVM.
|
|
services.chrony.enable = true;
|
|
services.chrony.servers = lib.mkForce [ ];
|
|
services.chrony.extraConfig = ''
|
|
refclock PHC /dev/ptp0 poll 2 prefer require refid KVM
|
|
makestep 0.1 3
|
|
'';
|
|
|
|
# Enable CockroachDB. In order to ensure that Chrony has performed its
|
|
# first synchronization at boot-time (which may take ~10 seconds) before
|
|
# starting CockroachDB, we block the ExecStartPre directive using the
|
|
# 'waitsync' command. This ensures Cockroach doesn't have its system time
|
|
# leap forward out of nowhere during startup/execution.
|
|
#
|
|
# Note that the default threshold for NTP-based skew in CockroachDB is
|
|
# ~500ms by default, so making sure it's started *after* accurate time
|
|
# synchronization is extremely important.
|
|
services.cockroachdb.enable = true;
|
|
services.cockroachdb.insecure = true;
|
|
services.cockroachdb.openPorts = true;
|
|
services.cockroachdb.locality = locality;
|
|
services.cockroachdb.listen.address = myAddr;
|
|
services.cockroachdb.join = lib.mkIf (joinNode != null) joinNode;
|
|
|
|
# Hold startup until Chrony has performed its first measurement (which
|
|
# will probably result in a full timeskip, thanks to makestep)
|
|
systemd.services.cockroachdb.preStart = ''
|
|
${pkgs.chrony}/bin/chronyc waitsync
|
|
'';
|
|
};
|
|
|
|
in import ./make-test-python.nix ({ pkgs, ...} : {
|
|
name = "cockroachdb";
|
|
meta.maintainers = with pkgs.stdenv.lib.maintainers;
|
|
[ thoughtpolice ];
|
|
|
|
nodes = {
|
|
node1 = makeNode "country=us,region=east,dc=1" "192.168.1.1" null;
|
|
node2 = makeNode "country=us,region=west,dc=2b" "192.168.1.2" "192.168.1.1";
|
|
node3 = makeNode "country=eu,region=west,dc=2" "192.168.1.3" "192.168.1.1";
|
|
};
|
|
|
|
# NOTE: All the nodes must start in order and you must NOT use startAll, because
|
|
# there's otherwise no way to guarantee that node1 will start before the others try
|
|
# to join it.
|
|
testScript = ''
|
|
for node in node1, node2, node3:
|
|
node.start()
|
|
node.wait_for_unit("cockroachdb")
|
|
node1.succeed(
|
|
"cockroach sql --host=192.168.1.1 --insecure -e 'SHOW ALL CLUSTER SETTINGS' 2>&1",
|
|
"cockroach workload init bank 'postgresql://root@192.168.1.1:26257?sslmode=disable'",
|
|
"cockroach workload run bank --duration=1m 'postgresql://root@192.168.1.1:26257?sslmode=disable'",
|
|
)
|
|
'';
|
|
})
|