mtail: cap memory

See merge request !35

Swap is taken into account too in cgroups v1, disable it.

Override LimitNOFILE in the instance-specific service override file.

In commit ae50318d some modified dashboards were exported using the
wrong method, which resulted in the classic DS_LOCALHOST errors.

Partial rollback of ae50318d where we mistakenly overwrote the wrong
dashboard.

To 90 days by default (configurable with prometheus_tsdb_retention).

Some of these systemd metrics have large cardinalities, so it's good
to prune them at collection time.

Part of optimization work for issue #51.

'null' is special.

The root route must not have matchers, create a child one instead.

Alerts of other severity levels are kept as advisory and not routed
anywhere.

Removes the instance-specific 'monitor' label set on alerts, so that
multiple Prometheus instances in a HA setup all create the same
alerts.

Previously we were also counting 4xx errors, which are not relevant
as a reliability indicator. We may eventually want to restrict this
further to 502/503 (for backend timeouts only).

Caused by a typo.

Part of the optimization work on issue #51.

This metric takes quite a bit of evaluation time due to the large
cardinality of node_systemd_unit_state, and is unused.

Part of the optimization work in issue #51.

Display nice HTTP error pages where possible

See merge request !34