IGNITE-28692 Add cache metrics for MDC-related checks

[sergey-chugunov-1985]

Thank you for submitting the pull request to the Apache Ignite.

In order to streamline the review of the contribution
we ask you to ensure the following steps have been taken:

The Contribution Checklist

• There is a single JIRA ticket related to the pull request.
• The web-link to the pull request is attached to the JIRA ticket.
• The JIRA ticket has the Patch Available state.
• The pull request body describes changes that have been made.
  The description explains WHAT and WHY was made instead of HOW.
• The pull request title is treated as the final commit message.
  The following pattern must be used: IGNITE-XXXX Change summary where XXXX - number of JIRA issue.
• A reviewer has been mentioned through the JIRA comments
  (see the Maintainers list)
• The pull request has been checked by the Teamcity Bot and
  the green visa attached to the JIRA ticket (see tab PR Check at TC.Bot - Instance 1 or TC.Bot - Instance 2)

Notes

• How to Contribute
• Coding abbreviation rules
• Coding Guidelines
• Apache Ignite Teamcity Bot

If you need any help, please email dev@ignite.apache.org or ask anу advice on http://asf.slack.com #ignite channel.

[anton-vinogradov]

Reviewed against IGNITE-28692. The two metrics are a reasonable direction, but a few things are worth addressing before merge — including one logic bug that the current tests don't catch and a per-exchange overhead that also hits non-MDC clusters. Details inline.

Higher level:

1. The JIRA lists three deliverables — a warning on cache start plus the two metrics. The warning-on-cache-start isn't part of this PR. Is it intentionally split out? It's arguably the most visible signal for a user who simply forgot the cache-level configuration; a passive boolean metric is unlikely to reach exactly that user.

2. IsCacheAffinityConfigurationMdcSafe is a heuristic over filter types, so it can't be a real guarantee in either direction: ClusterNodeAttributeColocatedBackupFilter is only MDC-safe when the cells are actually stretched across DCs (a deployment property, invisible from the filter), and a non-Rendezvous custom affinity is currently always reported as safe. Might be worth reflecting that in the metric description / Javadoc so Safe isn't read as a guarantee.

3. For the distribution metric, a boolean loses diagnostic value — a count of partitions without full DC coverage would also serve as a severity signal. Just a thought, not blocking.

[anton-vinogradov]

ClusterNodeAttributeAffinityBackupFilter places backups on nodes with a different value of the attribute, so with DATA_CENTER_ID it forces the backup into another DC — i.e. it is MDC-safe. But it's neither MdcAffinityBackupFilter nor ClusterNodeAttributeColocatedBackupFilter, so this if returns false for it. By the time we reach here such a filter is already guaranteed to contain DATA_CENTER_ID (otherwise we'd have returned at the check above), so it's safe to accept:

Suggested change

	if (!(filter instanceof MdcAffinityBackupFilter) && !(filter instanceof ClusterNodeAttributeColocatedBackupFilter))
	return false;
	if (!(filter instanceof MdcAffinityBackupFilter)
	&& !(filter instanceof ClusterNodeAttributeColocatedBackupFilter)
	&& !(filter instanceof ClusterNodeAttributeAffinityBackupFilter))
	return false;

This case isn't covered today: CACHE_WITH_MDC_SAFE_ATTRIBUTE_FILTER uses exactly this filter, but MdcCacheMetricsTest only checks its distribution metric, never IsCacheAffinityConfigurationMdcSafe — worth adding that assertion.

Minor follow-up for the multi-attribute case: getAttributeNames().contains(DATA_CENTER_ID) is necessary but not sufficient — with e.g. [DATA_CENTER_ID, RACK] the AND-semantics still allow a backup in the same DC but a different rack.

[anton-vinogradov]

numberOfDatacenters() returns -1 when DCs aren't configured, but that isn't handled here: for a non-MDC cluster we still iterate every partition of every non-replicated group on the exchange thread (allocating a stream per partition) and then discard the result — on every PME. An early return honours the -1 contract and skips the work for non-MDC clusters; it also avoids leaving the metric at its default true when the DC count is unknown:

Suggested change

	int numberOfDataCenters = top.numberOfDatacenters();
	int numberOfDataCenters = top.numberOfDatacenters();
	if (numberOfDataCenters < 2)
	return;

[anton-vinogradov]

These are written from the exchange thread (setMdcSafePartitionDistribution) and read by metric exporters / JMX from other threads. Without volatile there's no happens-before, so readers may observe a stale value — more likely to actually surface on weakly-ordered (ARM) CPUs. The other metric fields here are thread-safe (LongAdderMetric); these should be too:

Suggested change

	private Boolean affCfgMdcSafe;
	/** */
	private Boolean mdcSafePartDistrib;
	private volatile Boolean affCfgMdcSafe;
	/** */
	private volatile Boolean mdcSafePartDistrib;

[anton-vinogradov]

Inspecting only the first node's attributes makes this return -1 whenever the first baseline node happens to lack the DC attribute — even if every other node has it. That's exactly the partial-misconfiguration case this feature is meant to catch, and it would silently mark the whole cluster as safe. Counting across all nodes (ignoring nulls) is more robust:

Suggested change

	Collection<Map<String, Object>> allNodesAttrs = nodeMap.values();
	if (!allNodesAttrs.isEmpty() && allNodesAttrs.iterator().next().get(ATTR_DATA_CENTER_ID) != null)
	return (int)allNodesAttrs.stream().map(m -> m.get(ATTR_DATA_CENTER_ID)).distinct().count();
	return -1;
	long dcs = nodeMap.values().stream()
	.map(m -> m.get(ATTR_DATA_CENTER_ID))
	.filter(dc -> dc != null)
	.distinct()
	.count();
	return dcs == 0 ? -1 : (int)dcs;

[anton-vinogradov]

Copy-paste: this repeats the colocated assertion. It should null-check the third metric (currently cacheWithMdcSafeAttrFilterDistributionSafeMetric is never asserted non-null):

Suggested change

	assertNotNull(cacheWithColocatedFilterDistributionSafeMetric);
	assertNotNull(cacheWithMdcSafeAttrFilterDistributionSafeMetric);

[anton-vinogradov]

Accidental double-submit — please disregard this review; see my other review on this PR (the one with the inline comments).