Bug in MetadataMigrateToDataTiersRoutingService - _tier_preference could be incorrect #81633
Description
Elasticsearch version (bin/elasticsearch --version): 7.16.0
(but the problem was likely present since the introduction of this service in #73689).
Plugins installed: []
JVM version (java -version):
OS version (uname -a if on a Unix-like system):
Description of the problem including expected versus actual behavior:
MetadataMigrateToDataTiersRoutingServiceis assuming that if an index is configured with a_tier_preferencethen it is the right data tier index allocation (c.f code):
// look at the value, get the correct tiers config and update the settings
if (currentIndexSettings.keySet().contains(TIER_PREFERENCE)) {
newSettingsBuilder.remove(attributeBasedRoutingSettingName);
logger.debug("index [{}]: removed setting [{}]", indexName, attributeBasedRoutingSettingName);
} else {
// parse the custom attribute routing into the corresponding tier preference and configure it
String attributeValue = currentIndexSettings.get(attributeBasedRoutingSettingName);
String convertedTierPreference = convertAttributeValueToTierPreference(attributeValue);
Steps to reproduce:
We have seen in the field indices configured with the following settings:
- Example 1:
"routing" : {
"allocation" : {
"include" : {
"_tier_preference" : "data_hot"
},
"require" : {
"data" : "warm"
}
}
},
- Example 2:
"routing" : {
"allocation" : {
"include" : {
"_tier_preference" : "data_content"
},
"require" : {
"data" : "warm"
}
}
},
In both examples, these indices were actually in the warm phase (confirmed with the ILM explain API results). As a user, if I were to run the _ilm/migrate_to_data_tiers API when migrating from nodes types to nodes roles, this would result in an incorrect allocation. These indices will then be allocated to the hot and content and data tier - instead of the warm data tier.
The only way to recover from this is to identify the corresponding ILM phase for the affected indices and set the data tier index allocation accordingly. As a user, it is impossible to do so out of the box (i.e it requires some scripting to cross check information from the list of indices, index settings and ILM explain results of each index).