[IGNITE-20379] Concurrency issues with write intent tracking

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 3.0.0-beta2
• Labels: ignite-3
• Ignite Flags: Docs Required, Release Notes Required
• Assignee: Kirill Sizov
• Reporter: Kirill Sizov
• Reviewer: Alexander Lapin
• Created: 07/Sep/23 13:47
• Updated: 11/Sep/23 07:35
• Resolved: 11/Sep/23 07:35

Description

There are a few concurrency issues with the write intent tracking code

1. See PendingRows class. The value is a TreeSet, which is not a concurrent collection. Thus it's not safe to do the read this way:

  public Set<RowId> getPendingRowIds(UUID txId) {
        return txsPendingRowIds.getOrDefault(txId, EMPTY_SET);
    }

2. See StorageUpdateHandler.handleTransactionCleanup. We first get a collection of rows for a specific transaction from the map, then process the rows in the storage, then remove the whole transaction entry from the map:

    public void handleTransactionCleanup(UUID txId, boolean commit, @Nullable HybridTimestamp commitTimestamp) {
        Set<RowId> pendingRowIds = pendingRows.getPendingRowIds(txId);

        handleTransactionCleanup(pendingRowIds, commit, commitTimestamp, () -> pendingRows.removePendingRowIds(txId));
    }

If a parallel thread adds more write intents to the map, we'll silently lose them.

Issue Links

• is depended upon by: IGNITE-20041 Asynchronous writeIntent cleanup on write intent resolution
• links to: GitHub Pull Request #2562