Decision Guidance for Serverless Adoption

S. No.

Characteristic

Detail

Rationale

1

Short Process Running Time

Functions, applications or services that run for a finite and short time, typically CPU or minimal IO bound.

Functions, applications or services' running time affect two important factors:

Cost: Cost is based on execution time (in ms).

Concurrency: Processing time affects the throughput and concurrency. The process with a Long running time reduces the throughput. synchronous flows may fail due to time-outs. 

The following example explains the relation between concurrency, a process running time, and throughput.

Concurrency = 2, processing time is 400 ms, and throughput will be 5. If there are more than 5 requests, these will be queued up or throttled.

Simply put, Concurrency = Process Running Time x Request rate per second

The process running time should be short consistently to keep the cost and concurrency in check.

2

Quick Start up

Lightweight Applications, functions, or services with simple dependencies that use asynchronous execution with quick start-up time. Start-up time may include:

• Retrieve Source code
• Start the deployment form factor e.g., container
• Bootstrap any initialization as required

After successful bootstrap and initialization, applications, functions, or services can accept requests to execute

Typically; Start-up time influences two of serverless challenges 1. Cold Starts 2. Traffic Patterns. Long Start-up time affects the ability to provide consistent throughput. For a traffic pattern that matches a period of idle time (no requests) followed by a sudden spike or burst. During the spike of burst time, functions with significant cold start fail to serve the requests. Inconsistent traffic scenarios, a specified number of function instances are kept in the initialized or warm state by the platform anticipating some requests to be served. Start-up should not impact the required throughput or concurrency in any case, including different traffic scenarios.

3

Consistent and Predictable running time

Applications, functions, or services that run for consistent and predictable time irrespective of the way and source of the trigger. Variation or semantics of the source of trigger or requests should not impact the running time consistency or predictability.

Running time impacts the cost, concurrency (and hence throughput) as described in Short Process Running Time characteristic.

4

Consistent and Predictable Usage of Memory

Serverless platform requires the memory allocation choice as input. Services that use consistent and predictable memory, irrespective of event or a request payload that triggered these, are more suitable for serverless compute form factor than the workloads for which memory requirements vary dynamically.

Typically, the serverless platform takes the 'memory to be allocated' input and allocated Compute (CPU) based on it.  Hence it is important to select the right memory allocation carefully as it controls the allocation of resources like compute by platform and is considered as the key lever for serverless resource allocation.

5

Loose Coupling or Decoupled patterns such as Asynchronous, Reactive, Event Driven

The genesis of serverless advocates event-driven pattern that essentially provides loose coupling or decoupling. Leveraging the Reactive framework results in IO unlocking or unblocking and optimized use of computing.

The paradigm of Synchronous execution has two issues i.e., thread blocking and request/execution queueing

In hybrid cloud and distributed environments, this results in poor scalability. The scalability is enhanced significantly with an event-driven pattern that unlocks or unblocks the execution instances, and execution queues are managed by the event-triggering sources. As a result, scaling becomes more dynamic and efficacious.

6

Unpredictable spikes and bursts traffic patterns

Traffic patterns that cannot be predicted inherently comes with sporadic spikes or bursts. The premise of cloud-native implementation addresses this through elastic scalability making the services resilient and responsive in such situations.

Serverless compute services are automatically scaled in response to request demand by the platform. Serverless brings this unique differentiation. Without any auto-scaling rules and complexity, the serverless platform handles the request spike and burst, given the specified limit of concurrency and burst capacity. The platform can provide the required number of instances for serverless functions and services to address request spikes if the cold start issue is well managed.

7

Stateless principle of cloud native workloads

One of the key tenets of 12 factor Cloud Native application is Statelessness. Horizontal scaling helps improve performance by making the required number of instances available to serve the request traffic. Dependency for maintaining state or request affinity to a specific instance leads to ineffective scaling.

Due to the inherent characteristic of event-driven execution, serverless computing doesn't keep any previous context of request execution. Every instance will execute the request as if it is serving a new request. Other solutions like distributed cache or environment variables can be considered if any state or request context needs to be managed or shared.

8

Execution in parallel

Parallel execution threads process a service independently without any shared or global state or context.

Processes or Workloads that can be decomposed into multiple parallel independent functions are suitable for Serverless applications (e.g., processing a large set of structured and unstructured data, Big Data). These types of workloads allow execution of multiple processes as parallel functions.

9

Long Running Process that can be broken into small, short running processes

Often long-running process that doesn’t maintain a resilient process state can be decomposed into a short-running process that can be executed in parallel.

Moving to serverless architecture with parallelism and a scalable model for short-running processes will help in reducing sustained execution environment need resulting in lower TCO and improving agility and resiliency.

10

Minimal or no periods of inactivity

Applications, services, or workload that have the minimum number of requests served without idle time.

It is sometimes obvious to think that workloads that have a large number of inactivity periods and unpredictable spikes/bursts are suitable for serverless as the platform does not charge for idle time for serverless computing. However, performance is significantly impacted in such cases as the initialization, and bootstrapping context of instances are gone with the instances and cannot be reused as the platform removes these. Such cases are better suited for containers. A minimum consistent number of requests helps maintain a set of initialized and warm functions to serve any traffic.

11

Databases with unpredictable and erratic usage pattern

Databases that are implemented as serverless services can scale elastically

Capabilities: A) Databases as serverless services provided by platforms offer capacity-based costing where consumers are charged for the capacity consumed. B) Serverless Database services are elastic in a way that platform manages the underlying storage.

For high transaction throughput requirements, the performance provided by Serverless Database services is linear.

12

Quick validation of new ideas

Validating a new idea/hypothesis that requires quick PoC/PoV.

Adoption of serverless removes the need for provisioning and configuring any infrastructure. So the function and services can be built rapidly and tested.

13

Continuous data streams that are not bound

Data streams that are continuous and required to be processed by stream functions such as filters, maps aggregation etc.

Backend Services, along with FaaS providing processing hooks, can process continuous data streams that are unbound.

14

Processing of IIOT and general IoT Sensor Data

This also qualifies as Continuous data streams that are not bound.

Stream processing, IoT analytic & IoT protocol support e.g., MQTT, are provided by Cloud Service Providers.

15

Simplified Operation

Server maintenance is being viewed as an overhead for the organizations.

The serverless platform provided managed operation services that can reduce the burden of managing the server environment and simplify operation.