A Literature Review on Serverless Computing

Jinfeng Wen, Zhenpeng Chen, Xuanzhe LiuarXiv 2022 · OpenReview

Serverless computing, often synonymous with Function-as-a-Service (FaaS), has revolutionized cloud development by promising a "NoOps" experience. By allowing developers to deploy code without managing underlying infrastructure and offering a granular "pay-as-you-go" billing model, it has seen explosive adoption. However, this paradigm shift introduces significant technical hurdles. This paper provides a systematic map of the research landscape, analyzing 164 influential papers to build a 17-category taxonomy that addresses the field's most pressing challenges.

Defining the Serverless Landscape

At its core, serverless computing is composed of two main pillars: Function-as-a-Service (FaaS) and Backend-as-a-Service (BaaS). FaaS allows developers to write logic as discrete functions triggered by events, while BaaS provides managed services like databases, storage, and notifications that these functions interact with.

The architectural relationship between FaaS and BaaS in serverless platforms.

Despite the simplicity for the end-user, several "headaches" persist for researchers and practitioners:

  • The Cold Start: The latency penalty incurred when a platform must spin up a new container for an idle function.
  • Statelessness: Functions are short-lived and do not retain data between executions, complicating stateful applications like AI training.
  • Resource Management: The difficulty of choosing optimal memory and CPU settings to balance cost and performance.
  • Vendor Lock-in: The lack of standardization makes migrating between providers (e.g., AWS to Google Cloud) a non-trivial task.

Systematic Methodology

The authors conducted a Systematic Literature Review (SLR) following the Kitchenham guidelines to ensure a comprehensive and unbiased overview. Their process involved:

  1. Search & Filter: Querying seven major databases (including ACM and IEEE), yielding 398 initial papers.
  2. Refinement: Applying strict inclusion/exclusion criteria to narrow the selection to 118 high-quality peer-reviewed papers.
  3. Snowballing: Using reference tracking to identify an additional 46 relevant works, totaling 164 papers published between 2014 and 2022.
  4. Taxonomy Construction: Utilizing an "open coding" approach where researchers independently categorized papers into a hierarchical taxonomy of 11 root categories and 17 specific research directions.

Key Insights and the Research Taxonomy

The resulting taxonomy reveals that the research community is primarily obsessed with Performance (40.85%). Within this category, the "Cold Start" problem is the single most studied topic (34 papers), followed closely by function communication overhead.

Another significant area is Resource Management (13.41%). This highlights a "NoOps Paradox": while serverless aims to eliminate infrastructure management, developers still spend significant effort manually tuning memory settings to avoid over-provisioning costs. The review suggests a growing need for automated, autonomous resource allocation.

Comparing Commercial and Open-Source Platforms

Research is currently split between analyzing commercial giants and developing custom academic prototypes. Approximately 30% of studies utilize AWS Lambda due to its market dominance, while 31% build custom systems (like SAND or SOCK) to experiment with radical architectural changes.

Comparison of features across major commercial and open-source serverless platforms.

As shown in the comparison above, platforms vary significantly in their timeout limitations, memory allocation ranges, and observability tools. These discrepancies are a primary driver of vendor lock-in, as code optimized for AWS's 10GB memory limit or 900-second timeout may not function correctly on a platform with more restrictive constraints.

Future Frontiers

While the field is maturing, the authors identify several critical gaps that represent future research opportunities:

  • Stateful FaaS: Moving beyond simple web triggers to complex workflows requires finding ways for functions to share state without the latency of external databases.
  • Generic Conversion Tools: There is a lack of automated tools to decompose monolithic applications into optimized serverless functions.
  • The Computing Continuum: As serverless moves to the "Edge" (IoT devices), research must address data privacy and the extreme resource constraints of small hardware.
  • Testing and Debugging: Debugging distributed, event-driven applications remains a significant challenge, with a lack of mature, automated testing frameworks specifically designed for the FaaS lifecycle.

In conclusion, while serverless computing has solved many infrastructure problems, it has birthed a new generation of systems and software engineering challenges that will define cloud research for the coming decade.

Summary updated

BibTeX
@misc{wen2022literature,
  author = {Jinfeng Wen and Zhenpeng Chen and Xuanzhe Liu},
  title = {{A Literature Review on Serverless Computing}},
  year = {2022},
  url = {https://openreview.net/forum?id=JU46huxQZat},
}
Figure