For some reason (Debian bug?) the default rsyslog-rotate script was not
properly rotating rsyslog logfiles. Instead, explicitly call systemctl
kill -s HUP for this, using a full path.
Uses a much nicer CPU tuning configuration, leveraging systemd's
AllowedCPUs and CPUAffinity options within a set of slices (some
default, some custom).
Configuration is also greatly simplified versus the previous
implementation, simply asking for a number of CPUS for both the system
and OSDs, and calculating everything else that is required.
Also switches (back) to the v2 unified cgroup hierarchy by default as
required by the systemd AllowedCPUs directive.
This functionality simply did not work, with Libvirt continuing to dump
its processes into the root cset thus defeating the purpose entirely.
Just remove it, from some very initial testing it isn't worth the
headache.
Add model and serial numbers to the vendor, and put this on its own
line. Also use BASH for proper syntax formatting. Reformat the header to
be a more compact format.
Ensures that the pool default size/min size is set to something
reasonable for a single node (effective RAID-1) and replace teh default
CRUSH replicate_rule set for this situation with one choosing OSD
instead of host as the default.
The new CheckMK agent uses UID 998 (dynamic) for itself. This causes
ownership problems with the old logic of this check. Move instead to a
range, where the UIDs from 200-599 are reserved for administrators, and
check for this range explicitly. Also eliminates the exceptions for ceph
and 2000 from previous iterations.
1. Remove the obsolete pvc-vacuum script install.
2. Remove notifies when modifying configs; we do not want to restart the
daemons uncontrolled.
3. Add bootstrap check to package installs so they only happen on
bootstrap.
This ensures this part of the role, on re-runs, will *only* update
configs and not actually touch the running daemon. This makes it safe to
run before a oneshot/update-pvc-daemons.yml playbook run.
This was causing some confusing conflicts, so create a new fact called
"this_node" which is inventory_hostname.split('.')[0], i.e. the short
name, and use that everywhere instead of an FQDN or true inventory
hostname.
This is required on Debian 11 to use the cset tool, since the newer
systemd implementation of a unified cgroup hierarchy is not compatible
with the cset tool.
Ref for future use:
https://github.com/lpechacek/cpuset/issues/40
1. Detail the caveats and specific situations and ref the documentation
which will provide more details.
2. Always install the configs, but use /etc/default/ceph-osd-cpuset to
control if the script does anything or not (so, the "osd" cset set is
always active just not set in a special way.
1. Remove an explicit OSD journal size, especially such a small one (no
clue why I ever added that...)
2. Add max scrubs, disable scrub during recovery, and set scrub sleep.
3. Add max backfills, tune recovery sleep to 0 to prioritize recovery.
Allows an administrator to set CPU pinning with the cpuset tool for Ceph
OSDs, in situations where CPU contention with VMs or other system tasks
may be negatively affecting OSD performance. This is optional, advanced
tuning and is disabled by default.
Coupled with the removal of explicit --image-features flags to the RBD
command in PVC itself, this ensures that only the two features supported
on kernel 4.19 are enabled by default.
This has no functional change on Buster, but on Bullseye this overrides
the stupid socket-based activation shenanigans that the default unit
tries to do, as well as the breaking replacement of the
/etc/default/libvirt variable names.
Due to the requirement of Ceph to have all peer nodes tightly
synchronized with each other to come online, PVC nodes need a way to
synchronize to each other even in the absence of an external time
reference. This is especially prevalent if a set of nodes are left
offline for an extended period (>1-2 weeks), since their hardware clocks
will drift. If the resulting Internet connectivity is then dependent on
a VM, this will cause a catch-22 and the cluster will not properly
start.
This configuration will accomplish that - if no suitable >6 stratum
peers are found, the hosts will enter orphan mode. Since they are now
all configured as "peers" with each other, they will collectively decide
on one of them to become the source and sync to it. A local stratum 10
fudge is added so that at least one of the nodes can become this source.
While this is not an ideal use of NTP, it is by far the cleanest
solution to this problem, and does not impact normal functionality when
the two configured stratum-2 servers are reachable.
This was an artifact of a much, much older Ceph configuration I ran, and
is not relevant with newer Ceph versions like those used in PVC.
Performance testing with Nautilus and Bluestore reveals a minimal
performance hit, and using `jemalloc` prevents cache autotuning from
being effective, so remove it.
Adds this option based on the findings of
https://github.com/python-zk/kazoo/issues/630, whereby restores of >1MB
in size would fail. This is considered an unsafe option, but given our
usecase no actual znode should ever exceed this limit; this is purely
for the large transactions that come from a `pvc task restore` action to
an empty Zookeeper instance.
Joshua Boniface