Mars Hall created PIO-102: ----------------------------- Summary: ESEngineInstances `getAll` results out of order (Elasticsearch 5.x) Key: PIO-102 URL: https://issues.apache.org/jira/browse/PIO-102 Project: PredictionIO Issue Type: Bug Components: Core Affects Versions: 0.11.0-incubating Reporter: Mars Hall Assignee: Mars Hall
Using the new Elasticsearch 5.x REST storage client as the meta storage source (`PIO_STORAGE_REPOSITORIES_METADATA_SOURCE=ELASTICSEARCH` setup in conf/pio-env.sh), I found that once an engine has been trained a certain number of times, that the most recent engine instance is no longer retrieved. So, I tracked down where those Elasticsearch queries originate. In the original Elasticsearch 1.x storage client, [the "scroll" pagination responses are collected by *appending*|https://github.com/apache/incubator-predictionio/blob/release/0.11.0/storage/elasticsearch1/src/main/scala/org/apache/predictionio/data/storage/elasticsearch/ESUtils.scala#L44] them to one another. In the new Elasticsearch 5.x client, [the "scroll" responses are collected by *prepending*|https://github.com/apache/incubator-predictionio/blob/release/0.11.0/storage/elasticsearch/src/main/scala/org/apache/predictionio/data/storage/elasticsearch/ESUtils.scala#L152] them to one another. This out-of-order concatenation breaks [ESEngineInstances `getLatestCompleted`|https://github.com/apache/incubator-predictionio/blob/release/0.11.0/storage/elasticsearch/src/main/scala/org/apache/predictionio/data/storage/elasticsearch/ESEngineInstances.scala#L192] by erroneously replacing the head of the results with an older engine instance, when there are enough engine instances to overflow a single page of Elasticsearch hits. I've observed this buggy behavior after ten trainings, when enough engine instances are stored to trigger Elasticsearch's scroll feature. I'll be opening a pull request shortly with the super-simple fix. -- This message was sent by Atlassian JIRA (v6.4.14#64029)