使用Gremlin的Azure Cosmos DB：RequestRateTooLargeException导致数据库状态不一致问题

Let me start by restating your issue clearly, since it's a tricky one that contradicts expected behavior:

You executed a large Gremlin query designed to create a main vertex plus multiple child vertices connected via edges—all within the same logical partition (sharing the same PartitionKey). When hitting the RU limit (400 RU/s in your test), a RequestRateTooLargeException was thrown. Instead of the entire transaction rolling back as promised by Cosmos DB's ACID guarantees, you found the main vertex already persisted in the database. Retrying the query then triggers a ConflictException because the main vertex's ID is already in use.

Your reproduction query perfectly captures the scenario:

g.addV('MyVertex').property(id,'33748fd1-32ff-440f-a261-580dbde4eece').property(single,'PartitionKey','PartitionKey'). 
sideEffect(__.addE('HasMassMembers').to(__.addV('MassMembers').property(single,'PartitionKey','PartitionKey').property(single,'PropertyA','A').property(single,'PropertyB','A').property(single,'PropertyC','A').property(single,'PropertyD','A').property(single,'PropertyE','A').property(single,'PropertyF','A').property(single,'PropertyG','A').property(single,'PropertyH','A').property(single,'PropertyI','A').property(single,'PropertyJ','A').property(single,'PropertyK','A').property(single,'PropertyL','A').property(single,'PropertyM','A').property(single,'PropertyN','A').property(single,'PropertyO','A').property(single,'PropertyP','A').property(single,'PropertyQ','A').property(single,'PropertyR','A').property(single,'PropertyS','A').property(single,'PropertyT','A').property(single,'PropertyU','A').property(single,'PropertyV','A').property(single,'PropertyW','A').property(single,'PropertyX','A').property(single,'PropertyY','A').property(single,'PropertyZ','A'))). 
sideEffect(__.addE('HasMassMembers').to(__.addV('MassMembers').property(single,'PartitionKey','PartitionKey').property(single,'PropertyA','A').property(single,'PropertyB','A').property(single,'PropertyC','A').property(single,'PropertyD','A').property(single,'PropertyE','A').property(single,'PropertyF','A').property(single,'PropertyG','A').property(single,'PropertyH','A').property(single,'PropertyI','A').property(single,'PropertyJ','A').property(single,'PropertyK','A').property(single,'PropertyL','A').property(single,'PropertyM','A').property(single,'PropertyN','A').property(single,'PropertyO','A').property(single,'PropertyP','A').property(single,'PropertyQ','A').property(single,'PropertyR','A').property(single,'PropertyS','A').property(single,'PropertyT','A').property(single,'PropertyU','A').property(single,'PropertyV','A').property(single,'PropertyW','A').property(single,'PropertyX','A').property(single,'PropertyY','A').property(single,'PropertyZ','A'))). 
sideEffect(__.addE('HasMassMembers').to(__.addV('MassMembers').property(single,'PartitionKey','PartitionKey').property(single,'PropertyA','A').property(single,'PropertyB','A').property(single,'PropertyC','A').property(single,'PropertyD','A').property(single,'PropertyE','A').property(single,'PropertyF','A').property(single,'PropertyG','A').property(single,'PropertyH','A').property(single,'PropertyI','A').property(single,'PropertyJ','A').property(single,'PropertyK','A').property(single,'PropertyL','A').property(single,'PropertyM','A').property(single,'PropertyN','A').property(single,'PropertyO','A').property(single,'PropertyP','A').property(single,'PropertyQ','A').property(single,'PropertyR','A').property(single,'PropertyS','A').property(single,'PropertyT','A').property(single,'PropertyU','A').property(single,'PropertyV','A').property(single,'PropertyW','A').property(single,'PropertyX','A').property(single,'PropertyY','A').property(single,'PropertyZ','A'))). 
sideEffect(__.addE('HasMassMembers').to(__.addV('MassMembers').property(single,'PartitionKey','PartitionKey').property(single,'PropertyA','A').property(single,'PropertyB','A').property(single,'PropertyC','A').property(single,'PropertyD','A').property(single,'PropertyE','A').property(single,'PropertyF','A').property(single,'PropertyG','A').property(single,'PropertyH','A').property(single,'PropertyI','A').property(single,'PropertyJ','A').property(single,'PropertyK','A').property(single,'PropertyL','A').property(single,'PropertyM','A').property(single,'PropertyN','A').property(single,'PropertyO','A').property(single,'PropertyP','A').property(single,'PropertyQ','A').property(single,'PropertyR','A').property(single,'PropertyS','A').property(single,'PropertyT','A').property(single,'PropertyU','A').property(single,'PropertyV','A').property(single,'PropertyW','A').property(single,'PropertyX','A').property(single,'PropertyY','A').property(single,'PropertyZ','A'))). 
sideEffect(__.addE('HasMassMembers').to(__.addV('MassMembers').property(single,'PartitionKey','PartitionKey').property(single,'PropertyA','A').property(single,'PropertyB','A').property(single,'PropertyC','A').property(single,'PropertyD','A').property(single,'PropertyE','A').property(single,'PropertyF','A').property(single,'PropertyG','A').property(single,'PropertyH','A').property(single,'PropertyI','A').property(single,'PropertyJ','A').property(single,'PropertyK','A').property(single,'PropertyL','A').property(single,'PropertyM','A').property(single,'PropertyN','A').property(single,'PropertyO','A').property(single,'PropertyP','A').property(single,'PropertyQ','A').property(single,'PropertyR','A').property(single,'PropertyS','A').property(single,'PropertyT','A').property(single,'PropertyU','A').property(single,'PropertyV','A').property(single,'PropertyW','A').property(single,'PropertyX','A').property(single,'PropertyY','A').property(single,'PropertyZ','A'))). 
sideEffect(__.addE('HasMassMembers').to(__.addV('MassMembers').property(single,'PartitionKey','PartitionKey').property(single,'PropertyA','A').property(single,'PropertyB','A').property(single,'PropertyC','A').property(single,'PropertyD','A').property(single,'PropertyE','A').property(single,'PropertyF','A').property(single,'PropertyG','A').property(single,'PropertyH','A').property(single,'PropertyI','A').property(single,'PropertyJ','A').property(single,'PropertyK','A').property(single,'PropertyL','A').property(single,'PropertyM','A').property(single,'PropertyN','A').property(single,'PropertyO','A').property(single,'PropertyP','A').property(single,'PropertyQ','A').property(single,'PropertyR','A').property(single,'PropertyS','A').property(single,'PropertyT','A').property(single,'PropertyU','A').property(single,'PropertyV','A').property(single,'PropertyW','A').property(single,'PropertyX','A').property(single,'PropertyY','A').property(single,'PropertyZ','A'))). 
sideEffect(__.addE('HasMassMembers').to(__.addV('MassMembers').property(single,'PartitionKey','PartitionKey').property(single,'PropertyA','A').property(single,'PropertyB','A').property(single,'PropertyC','A').property(single,'PropertyD','A').property(single,'PropertyE','A').property(single,'PropertyF','A').property(single,'PropertyG','A').property(single,'PropertyH','A').property(single,'PropertyI','A').property(single,'PropertyJ','A').property(single,'PropertyK','A').property(single,'PropertyL','A').property(single,'PropertyM','A').property(single,'PropertyN','A').property(single,'PropertyO','A').property(single,'PropertyP','A').property(single,'PropertyQ','A').property(single,'PropertyR','A').property(single,'PropertyS','A').property(single,'PropertyT','A').property(single,'PropertyU','A').property(single,'PropertyV','A').property(single,'PropertyW','A').property(single,'PropertyX','A').property(single,'PropertyY','A').property(single,'PropertyZ','A'))). 
sideEffect(__.addE('HasMassMembers').to(__.addV('MassMembers').property(single,'PartitionKey','PartitionKey').property(single,'PropertyA','A').property(single,'PropertyB','A').property(single,'PropertyC','A').property(single,'PropertyD','A').property(single,'PropertyE','A').property(single,'PropertyF','A').property(single,'PropertyG','A').property(single,'PropertyH','A').property(single,'PropertyI','A').property(single,'PropertyJ','A').property(single,'PropertyK','A').property(single,'PropertyL','A').property(single,'PropertyM','A').property(single,'PropertyN','A').property(single,'PropertyO','A').property(single,'PropertyP','A').property(single,'PropertyQ','A').property(single,'PropertyR','A').property(single,'PropertyS','A').property(single,'PropertyT','A').property(single,'PropertyU','A').property(single,'PropertyV','A').property(single,'PropertyW','A').property(single,'PropertyX','A').property(single,'PropertyY','A').property(single,'PropertyZ','A'))). 
sideEffect(__.addE('HasMassMembers').to(__.addV('MassMembers').property(single,'PartitionKey','PartitionKey').property(single,'PropertyA','A').property(single,'PropertyB','A').property(single,'PropertyC','A').property(single,'PropertyD','A').property(single,'PropertyE','A').property(single,'PropertyF','A').property(single,'PropertyG','A').property(single,'PropertyH','A').property(single,'PropertyI','A').property(single,'PropertyJ','A').property(single,'PropertyK','A').property(single,'PropertyL','A').property(single,'PropertyM','A').property(single,'PropertyN','A').property(single,'PropertyO','A').property(single,'PropertyP','A').property(single,'PropertyQ','A').property(single,'PropertyR','A').property(single,'PropertyS','A').property(single,'PropertyT','A').property(single,'PropertyU','A').property(single,'PropertyV','A').property(single,'PropertyW','A').property(single,'PropertyX','A').property(single,'PropertyY','A').property(single,'PropertyZ','A'))). 
sideEffect(__.addE('HasMassMembers').to(__.addV('MassMembers').property(single,'PartitionKey','PartitionKey').property(single,'PropertyA','A').property(single,'PropertyB','A').property(single,'PropertyC','A').property(single,'PropertyD','A').property(single,'PropertyE','A').property(single,'PropertyF','A').property(single,'PropertyG','A').property(single,'PropertyH','A').property(single,'PropertyI','A').property(single,'PropertyJ','A').property(single,'PropertyK','A').property(single,'PropertyL','A').property(single,'PropertyM','A').property(single,'PropertyN','A').property(single,'PropertyO','A').property(single,'PropertyP','A').property(single,'PropertyQ','A').property(single,'PropertyR','A').property(single,'PropertyS','A').property(single,'PropertyT','A').property(single,'PropertyU','A').property(single,'PropertyV','A').property(single,'PropertyW','A').property(single,'PropertyX','A').property(single,'PropertyY','A').property(single,'PropertyZ','A')))

Analysis of the Issue #

First, let's confirm the expected behavior: Cosmos DB explicitly guarantees atomicity for all operations within a single logical partition in a single Gremlin query. That means every write operation (vertex creation, edge creation) in your query should either all succeed or all fail. Partial commits like you're seeing shouldn't happen if the system is working as documented.

Let's rule out obvious mistakes in your query:

• Partition scoping: All vertices share the same PartitionKey, so every operation is tied to one logical partition. No cross-partition operations here, which would break atomicity.
• Query syntax: Your use of sideEffect is correct—it's meant to execute auxiliary operations without altering the main traversal flow, and these should still be part of the same partition-scoped transaction. There's no syntax error that would explain partial commits.

Given this, the observed behavior is almost certainly a bug in the Cosmos DB Gremlin API. When a RequestRateTooLargeException is thrown mid-execution of a multi-operation transaction, the system isn't properly rolling back already completed operations in that transaction.

Recommendations #

Here's what you can do to address this:

• Isolate the issue: Try reducing the number of sideEffect clauses in your query and see if the partial commit still occurs. This helps confirm whether the problem is tied to transaction size or a more general edge case.
• Collect diagnostics: If you can consistently reproduce the issue, gather the query ID, failure timestamp, and any error details from Cosmos DB logs. This data will be critical for Azure support to diagnose the bug.
• Implement a workaround: For the short term, you can:
  • After a RequestRateTooLargeException, check if the main vertex exists. If it does, verify all child vertices and edges are present, then fill in any missing ones instead of retrying the full query.
  • If safe to do so, delete the partially created main vertex before retrying the full query to avoid ConflictException.
• Report the bug: Reach out to Azure support with your reproduction steps and diagnostic data. This is likely an unhandled edge case in the Gremlin engine's transaction handling that needs to be fixed.

内容的提问来源于stack exchange，提问作者Martin Demberger