Diagnose and troubleshoot slow requests in Azure Cosmos DB .NET SDK

APPLIES TO: NoSQL

In Azure Cosmos DB, you might notice slow requests. Delays can happen for multiple reasons, such as request throttling or the way your application is designed. This article explains the different root causes for this problem.

Request rate too large

Request throttling is the most common reason for slow requests. Azure Cosmos DB throttles requests if they exceed the allocated request units for the database or container. The SDK has built-in logic to retry these requests. The request rate too large troubleshooting article explains how to check if the requests are being throttled. The article also discusses how to scale your account to avoid these problems in the future.

Application design

When you design your application, follow the .NET SDK best practices for the best performance. If your application doesn't follow the SDK best practices, you might get slow or failed requests.

Consider the following when developing your application:

Metadata operations

If you need to verify that a database or container exists, don't do so by calling Create...IfNotExistsAsync or Read...Async before doing an item operation. The validation should only be done on application startup when it's necessary, if you expect them to be deleted. These metadata operations generate extra latency, have no service-level agreement (SLA), and have their own separate limitations. They don't scale like data operations.

Slow requests on bulk mode

Bulk mode is a throughput optimized mode meant for high data volume operations, not a latency optimized mode; it's meant to saturate the available throughput. If you're experiencing slow requests when using bulk mode make sure that:

  • Your application is compiled in Release configuration.
  • You aren't measuring latency while debugging the application (no debuggers attached).
  • The volume of operations is high, don't use bulk for less than 1000 operations. Your provisioned throughput dictates how many operations per second you can process, your goal with bulk would be to utilize as much of it as possible.
  • Monitor the container for throttling scenarios. If the container is getting heavily throttled it means the volume of data is larger than your provisioned throughput, you need to either scale up the container, or reduce the volume of data (maybe create smaller batches of data at a time).
  • You're correctly using the async/await pattern to process all concurrent Tasks and not blocking any async operation.

Capture diagnostics

All the responses in the SDK, including CosmosException, have a Diagnostics property. This property records all the information related to the single request, including if there were retries or any transient failures.

The diagnostics are returned as a string. The string changes with each version, as it's improved for troubleshooting different scenarios. With each version of the SDK, the string will have breaking changes to the formatting. Don't parse the string to avoid breaking changes. The following code sample shows how to read diagnostic logs by using the .NET SDK:

try
{
    ItemResponse<Book> response = await this.Container.CreateItemAsync<Book>(item: testItem);
    if (response.Diagnostics.GetClientElapsedTime() > ConfigurableSlowRequestTimeSpan)
    {
        // Log the response.Diagnostics.ToString() and add any additional info necessary to correlate to other logs 
    }
}
catch (CosmosException cosmosException)
{
    // Log the full exception including the stack trace with: cosmosException.ToString()
    
    // The Diagnostics can be logged separately if required with: cosmosException.Diagnostics.ToString()
}

// When using Stream APIs
ResponseMessage response = await this.Container.CreateItemStreamAsync(partitionKey, stream);
if (response.Diagnostics.GetClientElapsedTime() > ConfigurableSlowRequestTimeSpan || !response.IsSuccessStatusCode)
{
    // Log the diagnostics and add any additional info necessary to correlate to other logs with: response.Diagnostics.ToString()
}

Diagnostics in version 3.19 and later

The JSON structure has breaking changes with each version of the SDK. This makes it unsafe to be parsed. The JSON represents a tree structure of the request going through the SDK. The following sections cover a few key things to look at.

CPU history

High CPU utilization is the most common cause for slow requests. For optimal latency, CPU usage should be roughly 40 percent. Use 10 seconds as the interval to monitor maximum (not average) CPU utilization. CPU spikes are more common with cross-partition queries, where the requests might do multiple connections for a single query.

The timeouts include diagnostics, which contain the following, for example:

"systemHistory": [
{
"dateUtc": "2021-11-17T23:38:28.3115496Z",
"cpu": 16.731,
"memory": 9024120.000,
"threadInfo": {
"isThreadStarving": "False",
....
}

},
{
"dateUtc": "2021-11-17T23:38:38.3115496Z",
"cpu": 16.731,
"memory": 9024120.000,
"threadInfo": {
"isThreadStarving": "False",
....
}

},
...
]
  • If the cpu values are over 70 percent, the timeout is likely to be caused by CPU exhaustion. In this case, the solution is to investigate the source of the high CPU utilization and reduce it, or scale the machine to a larger resource size.
  • If the threadInfo/isThreadStarving nodes have True values, the cause is thread starvation. In this case, the solution is to investigate the source or sources of the thread starvation (potentially locked threads), or scale the machine or machines to a larger resource size.
  • If the dateUtc time between measurements isn't approximately 10 seconds, it also indicates contention on the thread pool. CPU is measured as an independent task that is enqueued in the thread pool every 10 seconds. If the time between measurements is longer, it indicates that the async tasks aren't able to be processed in a timely fashion. The most common scenario is when your application code is blocking calls over async code.

Solution

The client application that uses the SDK should be scaled up or out.

HttpResponseStats

HttpResponseStats are requests that go to the gateway. Even in direct mode, the SDK gets all the metadata information from the gateway.

If the request is slow, first verify that none of the previous suggestions yield the desired results. If it's still slow, different patterns point to different problems. The following table provides more details.

Number of requests Scenario Description
Single to all Request timeout or HttpRequestExceptions Points to SNAT port exhaustion, or a lack of resources on the machine to process the request in time.
Single or small percentage (SLA isn't violated) All A single or small percentage of slow requests can be caused by several different transient problems, and should be expected.
All All Points to a problem with the infrastructure or networking.
SLA violated No changes to application, and SLA dropped. Points to a problem with the Azure Cosmos DB service.
"HttpResponseStats": [
    {
        "StartTimeUTC": "2021-06-15T13:53:09.7961124Z",
        "EndTimeUTC": "2021-06-15T13:53:09.7961127Z",
        "RequestUri": "https://127.0.0.1:8081/dbs/347a8e44-a550-493e-88ee-29a19c070ecc/colls/4f72e752-fa91-455a-82c1-bf253a5a3c4e",
        "ResourceType": "Collection",
        "HttpMethod": "GET",
        "ActivityId": "e16e98ec-f2e3-430c-b9e9-7d99e58a4f72",
        "StatusCode": "OK"
    }
]

StoreResult

StoreResult represents a single request to Azure Cosmos DB, by using direct mode with the TCP protocol.

If it's still slow, different patterns point to different problems. The following table provides more details.

Number of requests Scenario Description
Single to all StoreResult contains TransportException Points to SNAT port exhaustion, or a lack of resources on the machine to process the request in time.
Single or small percentage (SLA isn't violated) All A single or small percentage of slow requests can be caused by several different transient problems, and should be expected.
All All A problem with the infrastructure or networking.
SLA violated Requests contain multiple failure error codes, like 410 Points to a problem with the Azure Cosmos DB service or the client machine.
SLA violated StorePhysicalAddress are the same, with no failure status code. Likely a problem with Azure Cosmos DB.
SLA violated StorePhysicalAddress have the same partition ID, but different replica IDs, with no failure status code. Likely a problem with Azure Cosmos DB.
SLA violated StorePhysicalAddress is random, with no failure status code. Points to a problem with the machine.

For multiple store results for a single request, be aware of the following:

  • Strong consistency and bounded staleness consistency always have at least two store results.
  • Check the status code of each StoreResult. The SDK retries automatically on multiple different transient failures. The SDK is constantly improved to cover more scenarios.

RequestTimeline

Show the time for the different stages of sending and receiving a request in the transport layer.

  • ChannelAcquisitionStarted: The time to get or create a new connection. Connections can be created for numerous reasons such as: The previous connection was closed due to inactivity using CosmosClientOptions.IdleTcpConnectionTimeout, the volume of concurrent requests exceeds the CosmosClientOptions.MaxRequestsPerTcpConnection, the connection was closed due to a network error, or the application is not following the Singleton pattern and new instances are constantly created. Once a connection is established, it is reused for subsequent requests, so this should not impact P99 latency unless the previously mentioned issues are happening.
  • Pipelined: The time spent writing the request into the TCP socket. Request can only be written on a TCP socket one at a time, a large value indicates a bottleneck on the TCP socket which is commonly associated with locked threads by the application code or large requests size.
  • Transit time: The time spent on the network after the request was written on the TCP socket. Compare this number to the BELatencyInMs. If BELatencyInMs is small, then the time was spent on the network, and not on the Azure Cosmos DB service. If the request failed with a timeout, it indicates how long the client waited without response, and the source is network latency.
  • Received: The time between the response was received and processed by the SDK. A large value is normally caused by a thread starvation or locked threads.

ServiceEndpointStatistics

Information about a particular backend server. The SDK can open multiple connections to a single backend server depending upon the number of pending requests and the MaxConcurrentRequestsPerConnection.

  • inflightRequests The number of pending requests to a backend server (maybe from different partitions). A high number may lead to more traffic and higher latencies.
  • openConnections is the total Number of connections open to a single backend server. This can be useful to show SNAT port exhaustion if this number is very high.

ConnectionStatistics

Information about the particular connection (new or old) the request get's assigned to.

  • waitforConnectionInit: The current request was waiting for new connection initialization to complete. This will lead to higher latencies.
  • callsPendingReceive: Number of calls that was pending receive before this call was sent. A high number can show us that there were many calls before this call and it may lead to higher latencies. If this number is high it points to a head of line blocking issue possibly caused by another request like query or feed operation that is taking a long time to process. Try lowering the CosmosClientOptions.MaxRequestsPerTcpConnection to increase the number of channels.
  • LastSentTime: Time of last request that was sent to this server. This along with LastReceivedTime can be used to see connectivity or endpoint issues. For example if there are many receive timeouts, Sent time will be much larger than the Receive time.
  • lastReceive: Time of last request that was received from this server
  • lastSendAttempt: Time of the last send attempt

Request and response sizes

  • requestSizeInBytes: The total size of the request sent to Azure Cosmos DB
  • responseMetadataSizeInBytes: The size of headers returned from Azure Cosmos DB
  • responseBodySizeInBytes: The size of content returned from Azure Cosmos DB
"StoreResult": {
    "ActivityId": "bab6ade1-b8de-407f-b89d-fa2138a91284",
    "StatusCode": "Ok",
    "SubStatusCode": "Unknown",
    "LSN": 453362,
    "PartitionKeyRangeId": "1",
    "GlobalCommittedLSN": 0,
    "ItemLSN": 453358,
    "UsingLocalLSN": true,
    "QuorumAckedLSN": -1,
    "SessionToken": "-1#453362",
    "CurrentWriteQuorum": -1,
    "CurrentReplicaSetSize": -1,
    "NumberOfReadRegions": 0,
    "IsValid": true,
    "StorePhysicalAddress": "rntbd://127.0.0.1:10253/apps/DocDbApp/services/DocDbServer92/partitions/a4cb49a8-38c8-11e6-8106-8cdcd42c33be/replicas/1s/",
    "RequestCharge": 1,
    "RetryAfterInMs": null,
    "BELatencyInMs": "0.304",
    "transportRequestTimeline": {
        "requestTimeline": [
            {
                "event": "Created",
                "startTimeUtc": "2022-05-25T12:03:36.3081190Z",
                "durationInMs": 0.0024
            },
            {
                "event": "ChannelAcquisitionStarted",
                "startTimeUtc": "2022-05-25T12:03:36.3081214Z",
                "durationInMs": 0.0132
            },
            {
                "event": "Pipelined",
                "startTimeUtc": "2022-05-25T12:03:36.3081346Z",
                "durationInMs": 0.0865
            },
            {
                "event": "Transit Time",
                "startTimeUtc": "2022-05-25T12:03:36.3082211Z",
                "durationInMs": 1.3324
            },
            {
                "event": "Received",
                "startTimeUtc": "2022-05-25T12:03:36.3095535Z",
                "durationInMs": 12.6128
            },
            {
                "event": "Completed",
                "startTimeUtc": "2022-05-25T12:03:36.8621663Z",
                "durationInMs": 0
            }
        ],
        "serviceEndpointStats": {
            "inflightRequests": 1,
            "openConnections": 1
        },
        "connectionStats": {
            "waitforConnectionInit": "False",
            "callsPendingReceive": 0,
            "lastSendAttempt": "2022-05-25T12:03:34.0222760Z",
            "lastSend": "2022-05-25T12:03:34.0223280Z",
            "lastReceive": "2022-05-25T12:03:34.0257728Z"
        },
        "requestSizeInBytes": 447,
        "responseMetadataSizeInBytes": 438,
        "responseBodySizeInBytes": 604
    },
    "TransportException": null
}

Failure rate violates the Azure Cosmos DB SLA

Contact Azure support.

Next steps