Friday, July 13, 2012

Architecture Review - Scalability

Scalability is an important quality attribute of any system, be it hardware or software. But in most cases, the need for a scalability check or review is felt only when certain signs of scalability problems show up. Typically, the following are such signs that call for a scalability review of an existing application.
  • When changes requested on certain subsystems are turned down by the development team(s) citing that it is a complex subsystem and any change to it might call for huge efforts in terms of regression testing or else, it could lead to a bigger impact on the whole system. This indicates that there are certain components or sub systems which prevent the system from scaling. 
  • Months after production usage, the application performance gradually slows down and there is a tendency to accept the performance slow down or to pump in more hardware to compensate the slowdown. This is again is an important sign that the application is not scaling to take on the ever growing user base and transaction volumes.
There could be more signs that could indicate that there are scalability issues within the application. It is unfortunate that scalability reviews are not done in the initial design phase, so that these post production troubles won’t show up. While reviewing an existing application for potential scalability issues may be easy, the solutions for addressing those may not be really easy. That could be because of the underlying design & architecture of the application and its inter-dependencies with other systems in use. Let us examine certain important aspects to look into to spot potential scalability problems.

Distributed architecture: While distributed design is likely to improve performance, it could lead to scalability issues when one or more components or sub system relies on the local resources. Another reason for this to be reviewed with care is an ill designed system may call for too much of communication across physical and logical boundaries of various subsystems and would rely more on the communication infrastructure.

Component interaction: Examine how the components or subsystems are designed to interact with each other and how closely they are positioned. Too much of component interactions could lead to network congestion and also result in very high latencies which results in performance scalability issues later on when the usage increases. Measure the payload and the latency of such inter component interactions and isolate the components that need redesign. Keeping the data and the behaviour closer will reduce the interactions across boundaries and as a result keeps the latency under check.

Resource contentions: Look for potential limitations on the hardware or software resources used by the application. For instance, if the application produces huge amounts of log data, on the same disk where its transaction data is stored, the write requests may encounter resource contentions. Similarly, how fast the data files grow and how does the disk subsystems support such growth. Possible solutions for such issues are using resource pooling, message queues or such other asynchronous mechanisms.

Remote Communications: It would always be beneficial to limit the remote calls to the minimum or else, too many remote calls may expose the system too much on the reliability and availability of the communication infrastructure. Ensure required validations are performed ahead of the remote calls, so that unnecessary remote calls are avoided. Where possible, the remote calls should be stateless and asynchronous. Synchronous calls may hold up the communication channels and associated resources for longer period which could be the potential cause for performance and scalability issues. Use of message queues may help in decoupling subsystems from being held up for synchronous responses.

Cache Management: While use of Cache can help achieve better performance, it could also prevent the application from scaling in a load balanced environment, unless a distributed caching mechanism is designed and used.

State Management: Look for how the state of the persistent objects is being managed. Stateless objects scale better than the stateful objects. Distributed state management is the solution to address the state management issue in a load balanced environment. Always prefer stateless components or services as this will perform well and at the same time scale well.

Here are some of the best practices that help achieve high scalability
  • Prefer stateless asynchronous communications as this will free up resources considerably and supports load balancing.
  •  Design the application into multiple fault isolated subsystems with ability of being deployed on different hardware environments (or isolated application pools), so that faults in one subsystem does not impact the other subsystems. This partition can be either by service categories or by customer segments.
  •  Use distributed cache solutions, so that the cached data is available on multiple clustered environments.
  •  Use distributed databases with appropriate replication so that loads can be distributed.
  •  Do not depend too much on the specific capabilities of the RDBMS, as this might couple the application tightly to one vendor’s RDBMS. High degree of scalability can be achieved by keeping the business logic outside of the RDBMS.
  •  Spot the potential scalability issues early on by performing design reviews during development and by performing periodic load and performance tests.
  •  Do not ignore the capacity planning activity early during the pre-project phase, as it could significantly impact the application usage in production over a period of time. Also be aware of the data growth rates and have a road map to support the ever increasing data and volume growth.
  • Do not ignore the root cause analysis as many times when developers roll in a fix for a defect, they are not fixing the root cause, which could come back later as a scalability bottleneck.
Update:
Also read this MSDN Library article which lists down five key considerations for a scalable design.