Description
There is a retry logic for API getAllChildrenPaginated() when there is deletion in the transaction children(children created by a multi transaction operation) and the new page's start minCzxidOffset shifts. Current implementation uses the last child as a flag to represent if there is offset shifting.
|
} else { |
|
// If a child with the same czxid is returned in the previous page, and then deleted |
|
// on the server, the children not in the previous page but offset is less than czxidOffset |
|
// would be missed in the next page. Use the last child in the previous page to determine whether or not |
|
// the deletion case happens. If it happens, retry fetching the page starting from offset 0. |
|
String lastChild = childrenPage.isEmpty() ? "" : childrenPage.get(childrenPage.size() - 1); |
|
childrenPage = getChildren(path, watcher, Integer.MAX_VALUE, minCzxid, |
|
nextPage.getMinCzxidOffset() - 1, null, nextPage); |
|
if (!lastChild.equals(childrenPage.get(0))) { |
|
if (retryCount++ >= 5) { |
|
LOG.warn("Failed to fetch children pages because of znode deletion and retries exceeding 5"); |
|
throw KeeperException.create(Code.OPERATIONTIMEOUT); |
|
} |
|
LOG.info("Due to znode deletion, retry fetching children from offset 0 for minCzxid={}, retries={}", |
|
minCzxid, retryCount); |
|
childrenPage = getChildren(path, watcher, Integer.MAX_VALUE, minCzxid, 0, null, nextPage); |
|
} |
|
} |
|
allChildrenSet.addAll(childrenPage); |
|
} while (nextPage.getMinCzxid() != ZooDefs.GetChildrenPaginated.lastPageMinCzxid); |
We want to add a unit test for the code logic. As we could not simulate the read/write race condition on the server, we will need to mock the states in the zookeeper api for the uni test.
Related to #8
