DynamoDB unmarshalling slow

[michelemottini]

Describe the bug

Unmarshalling DynamoDB GetItem / BatchGetItem response is slow - ~700 msec for a 1.7 MB response

Regression Issue

• Select this option if this issue appears to be a regression.

Expected Behavior

Unmarshalling time << of the HTTP request time

Current Behavior

Unmarshalling time >> of the HTTP request time

Reproduction Steps

Copy the following files in a S3 folder:

djk7enshnmzkphkldeszvgzole.json.gz
ikhj4i2zsi4ptfwkszjkizzvly.json.gz
jmknnpggim2u3fd3lknw5njlsq.json.gz
m5lnl37q7a7z7hvykktyqt2mn4.json.gz

Use the DynamoDB 'Import from S3' function on that folder to create a table

• Table name dbperftest
• Data format: DynamoDB JSON compressed
• Partition key: @rid (string)
  -No sort key
• Capacity mode: on demand
• No indexes, default encryption at rest

Download, compile and run

DBPerfTestConsole.zip

• it expects AWS profile and region in the AWS_PROFILE and AWS_REGION environment variables (SSO is fine)
• it reads data from the dbperftest table in various ways - main one being reading 100 items using a single BatchGetItemCall - whose statistics are:

CredentialsRequestTime = 00:00:00.0000269; RequestSigningTime = 00:00:00.0003347; HttpRequestTime = 00:00:00.2243586; ResponseUnmarshallTime = 00:00:00.7423187; ResponseProcessingTime = 00:00:00.7452110; ClientExecuteTime = 00:00:00.9727216;

with an unmarshall time that is more than three times the HTTP request time

Possible Solution

No response

Additional Information/Context

No response

AWS .NET SDK and/or Package version used

AWSSDK.DynamoDBv2 3.7.403.3

Targeted .NET Platform

.NET Core 8

Operating System and version

Windows 11, Amazon Linux

[normj]

We are currently working on a V4 of the SDK. As part of that V4 we hoping to switch the SDK's JSON parsing to System.Text.Json from the its internal usage of a JSON library called LitJson. We aren't far enough yet with the change to measure the performance difference but we will use your sample code for measure the difference.

[ashishdhingra]

Needs review with the team as v4 backlog item.

[michelemottini]

Does this mean that this is not going to be fixed until V4? . . . and when would V4 be available?

[dscpinheiro]

We're not planning to change the serialization in V3 as that'd be too much of a breaking change. For V4, you can track the progress here: #3362

See also Norm's latest comment about V4's timeline: #3362 (comment)

[michelemottini]

I see . . .

I tried with the v4 preview and the unmarshalling time is better but still not great:

HttpRequestTime = 00:00:00.1873335; ResponseUnmarshallTime = 00:00:00.3164829; ResponseProcessingTime = 00:00:00.3181069; ClientExecuteTime = 00:00:00.5074811;

(it is 2 times slower than the raw DynamoDB call instead of 4+ . . . )

[normj]

Keep in mind that HttpRequestTime is the time to first byte but the full response is read while doing the unmarshalling. So the numbers are a bit misleading what is exactly the "HTTP" part and the "Unmarshalling" part.

[michelemottini]

Ah I see, thanks