[ https://issues.apache.org/jira/browse/BEAM-260?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15275128#comment-15275128 ]
Aljoscha Krettek commented on BEAM-260: --------------------------------------- I thought about this as well while working on the Flink Streaming side input support. Could it be enough to have something like {{WindowFn.getSideInputCleanupTime(BoundedWindow)}} that tells you when you can GC a side input window based on the main-input watermark. This would be called on the WindowFn of the side input, since it knows how the main-input windows are mapped to side inputs. > Know the getSideInputWindow upper bound so can gc side input state > ------------------------------------------------------------------ > > Key: BEAM-260 > URL: https://issues.apache.org/jira/browse/BEAM-260 > Project: Beam > Issue Type: Bug > Components: beam-model > Reporter: Mark Shields > Assignee: Frances Perry > > We currently have no static knowledge about the getSideInputWindow function, > and runners are thus forced to hold on to all side input state / elements in > case a future element reaches back into an earlier side input element. > Maybe we need an upper bound on lag from current to result of > getSideInputWindow so we can have a progressing gc horizon as we do for GKB > window state. -- This message was sent by Atlassian JIRA (v6.3.4#6332)