Time based Expiration/Generation of Records/Events/Messages for Actions

Question

I am creating a script for Task Assignment to multiple users. A task is assigned to a user, if it is not accepted in the next 30 minutes, I need to re-assign to another user.

Actually, I used DynamoDB for this, where every task assignment is attached to TTL after 30 minutes. When TTL expired I handled through Stream and check if it is accepted. If not, I was re-assigning and creating a new entry in the table with 30 minutes TTL.

I missed one concept of TTL that it is not expiring items in real-time and could take up to 48 hours.

Is there any other smart way to handle such use cases, currently I implemented this as -

add an index with TTL as sort key and event_type = Task as a partition key
query all the records every minute where TTL less than the current epoch, and batch deleting those records.

There are two challenges here -

First, as event_type is constant across the table, all data is hitting one partition of the index, which is not good for big volumes
it's a poll mechanism where I actually need to scan through all records, this is also not a scalable solution

I am exploring if we can do it smartly by a push mechanism. Any pointers or help to solve this use case?

score 2 · Answer 1 · answered Oct 29 '20 at 11:38

2

I think cron jobs checking your database is a must. You won't escape that.

What you can do is add Global Secondary Index with your expiration date. Then you can use query instead of scan on sorted values.

answered Oct 29 '20 at 11:38

karjan

936
1
7
17

And only project the attributes you need to make this decision to keep the GSI small and inexpensive as possible. – NoSQLKnowHow Oct 29 '20 at 18:30
Yes, I am doing the same as of now. Running a Lambda every minute to query the records less than the current epoch. As we cant use < or > operators in partition key query, hence put them into sort key. I can put expiration epoch as the partition key but masked to minute level and query that to optimize and scale it further. I am looking if we can do it on the seconds level and something in a smarter way. – Pranay Agarwal Nov 02 '20 at 12:48
1

If you need such accuracy, I'd go with step functions. Yan Cui wrote a great article about it https://theburningmonk.com/2019/06/step-functions-as-an-ad-hoc-scheduling-mechanism/ – karjan Nov 02 '20 at 16:13

score 0 · Answer 2 · answered Nov 01 '20 at 01:36

Consider using temporal.io. It allows modeling your business logic directly in code. Here is how it would be done in Java:

@WorkflowInterface
public interface TaskAssignment {
    @WorkflowMethod
    void assignTask();

    @SignalMethod
    void reportAccepted(String userId);
}

@ActivityInterface
public interface TaskAssignmentActivities {

  String pickAvailableUser();

  void assignTask(String userId);
}

public class TaskAssignmentImpl implements TaskAssignment {

  private static final Duration TASK_ASSIGNMENT_TIMEOUT = Duration.ofHours(1);
  private static final int MAX_ASSIGNMENT_ATTEMPTS = 10;

  private ActivityOptions options =
      ActivityOptions.newBuilder().setStartToCloseTimeout(Duration.ofSeconds(10)).build();

  private final TaskAssignmentActivities activities =
      Workflow.newActivityStub(TaskAssignmentActivities.class, options);

  private String userId;

  @Override
  public void assignTask() {
    for (int i = 0; i < MAX_ASSIGNMENT_ATTEMPTS; i++) {
      String userId = activities.pickAvailableUser();
      activities.assignTask(userId);
      // Block up to TASK_ASSIGNMENT_TIMEOUT or user equality condition is satisfied.
      Workflow.await(TASK_ASSIGNMENT_TIMEOUT, () -> userId.equals(this.userId));
      if (userId.equals(this.userId)) {
        break;
      }
    }
  }

  @Override
  public void reportAccepted(String userId) {
    this.userId = userId;
  }
}

This looks like normal code, but Temporal makes it fully fault-tolerant. So if your processes are restarted the state of the computation is fully preserved including local variables and thread stacks. So no need to talk to a database or a queue and use cron jobs.

Time based Expiration/Generation of Records/Events/Messages for Actions

2 Answers2