Github user arunmahadevan commented on the issue: https://github.com/apache/storm/pull/2218 This builds on top of the existing state checkpointing mechanism (documented here - https://github.com/apache/storm/blob/master/docs/State-checkpointing.md). Theres nothing extra added to the underlying checkpointing mechanism itself and its pretty straightforward. The tuples in window (think a FIFO queue) are split into multiple partitions so that they are more manageable and can be distributed/sharded via the underlying key-value state (redis/hbase etc). The modified partitions are saved during a checkpoint. During iteration the partition are loaded on demand from the underlying state backend as they are accessed. A subset of the most recently used partitions are cached in memory (backed by a guava Loading Cache). During checkpoint, the following are saved : 1. Any modified or newly created window partitions. 2. Any state needed to recover the Trigger/Eviction policies. 3. State thats exposed to the user where the user may have saved some values. Since the KV state does not guarantee any specific ordering of the keys during iteration, a separate structure is maintained to store the ordered partition Ids which is used during iteration to retrieve the partitions in order. This is also saved during the checkpoint. The above mechanism kicks in only if user choses to use the windowed state persistence, otherwise the current behavior (keeping the tuples in an in-memory queue) is retained. I would like to capture the design details in the PR/Documentation itself as far as possible than adding it in separate docs which tend to get lost.
--- If your project is set up for it, you can reply to this email and have your reply appear on GitHub as well. If your project does not have this feature enabled and wishes so, or if the feature is enabled but not working, please contact infrastructure at infrastruct...@apache.org or file a JIRA ticket with INFRA. ---