Ever wondered how does 1-click-buy works on Amazon? How does an e-commerce platform show the status of your order after the order is placed? What happens when you cancel your order right after you place an order, or after your item is shipped, or even delivered? How is all the activity related to an order tied to just one order Id? This blog will try to tackle such system design challenges and layout key insights on designing a workflow system.
In a distributed software system, it is often required to achieve some goals with a combination of multiple operations, such as downstream calls, database transactions, and authentications. A workflow engine is an orchestration system that will help you hand out worker processes and manage the central state. It is possible to encounter failure, network glitch, or timeout, or the host running the software can fail. Therefore, a reliable workflow engine design is essential to the smooth performance of the overall software system.
I will highlight the design considerations of a good workflow engine in this blog from the lessons I’ve learned in the past while working on designing highly scalable systems.
Simple Example: Payment workflow
Given the distributed nature of the software system, the workflow engines face a series of challenges:
A workflow is a finite, ordered series of actions transforming the start state to the end state. It can be modeled as a directed acyclic graph with multiple paths between two special states: the initial state and the end state. Each state is a node, and each action is an edge. There can be multiple actions given some conditions for each state, but only one action can be taken on a state, and the choice of action is well-defined. The initial and end state of a workflow is well defined, and the consequence (“change set”) of each action is well-defined. Therefore, each workflow state is well-defined.
The state of a workflow is the state of a collection of resources, database entries, or states in remote systems. An action in a workflow is a dependencies API call, a database transaction, etc., that changes at least one aspect of the state of the workflow, e.g., payment is completed. Item is shipped etc.
Following design principles should be considered to address the challenges:
A workflow engine can be divided into the following functional units.
Workflow workers perform actions and state transitions. There are a couple of ways of designing workflow workers (Let’s call them activity).
There are both advantages and disadvantages to all techniques. In the dedicated worker approach, workflow state is maintained by a single worker, eliminating the need of passing state from one worker to the next. All the actions of this workflow must be defined in a single place, which can be good or bad depending on the situation. However, this process does not allow batch processing, and thus a workflow scheduler that distributes workflows to workers is required.
In the assembly line approach, workflow actions are spread to multiple workers, reducing the readability of the code. It requires passing workflow state between workers but allows batch processing.
It is recommended to model complicated, less frequent workflows with the dedicated worker approach and define many simple workflows using the assembly line approach.
You will always have some activities which run faster than other activities in a workflow. If you get n number of requests, then a particular activity from all the workflows may be completed, but the remaining activities are just piling up in the queue. To ensure workflows are organized desirably, prioritize on-the-fly tasks over new tasks, or distribute workflow tasks to multiple hosts running the workflow engine to avoid conflict, a workflow scheduler is helpful to manage workers so that workflows can be processed in the desired way.
The scheduler should adjust the speed of each worker so that it is limited to the slowest one to ensure no accumulation of work items becomes operational debt. The number of concurrent workers should be limited so that no dependency is overwhelmed just because lightweight activities are completed quickly.
Capacity planning and Rate Limiter
There is usually a limit of capacity for any dependencies. This is often defined as allowed Transaction per Second, or TPS, from a given service. Thus, the workflow engine should implement a rate limiter at the client-side to efficiently use the fleet capacity and network bandwidth.
The implementation of a Rate Limiter depends on the design of the worker. If workers are designed as threads, then the threads must be paused when allowed TPS is exceeded. If workers are designed as event handlers, API calls or database transactions should be registered as events scheduled with proper internal between them not to exceed the TPS requirement. The workflow engine should handle throttled requests with an appropriate back-off and retry mechanism.
Activities in a Workflow blocked on state change of asynchronous operations must be handled efficiently. For example, if payment confirmation notification from the bank is delayed, the order shouldn’t be confirmed. This use case must be handled using a wait queue. Similarly, workflows that are waiting for an asynchronous operation to complete should be handled separately. The wait queue in the workflow engine should gather pending activities and perform state polling at a much slower rate than the workers. Wait queue should handle notification of completion. The wait queue should remove messages from the queue as the related activities are completed and notify the worker.
Subscribe to get well designed content on data structure and algorithms, machine learning, system design, object orientd programming and math.
©2023 Code Algorithms Pvt. Ltd.
All rights reserved.