In July 2016, the world of gaming was taken over by a new phenomenon – Pokémon Go. Within a matter of days "augmented reality" became mainstream and the app, which was launched mainly in the US and Australia, overtook Tinder and Twitter in the total number of downloads. Pokémon Go surpassed the wildest expectations of its creators, Niantic Labs, and then some.
With popularity comes scale, and with scale comes an overload of requests to the gaming servers. If you are not prepared enough, requests fail and users are frustrated. Frustrated with Pokémon Go crashes, laymen were talking about server status and memes were being created and circulated on social networks. Overnight, websites spun up just to report if the game was up or down in different countries. Being closely related to the APM space, my head was drawing up various ways in which Pokémon Go was perhaps addressing the issue and what monitoring they had to put in place to retain their popularity. Here is my list of probable solutions Pokémon Go could employ to improve the experience for their users and avid fans:
1. Synthetic monitoring
The first and most important question: Is the application up or down, and can users login from around the world?
A game this popular would need to ensure a five nine availability and high Apdex score. With synthetic end-user monitoring, simulated tests can be run from around the world to check for availability and response time as often every few seconds. The simulations can allow you to login to the app and interact with the app as a gamer would.
For example if a user is catching a Pokémon, he makes a HTTP request to an API “catchPokemon” with a set of parameters. Continuously checking if these HTTP requests return a valid response code within a reasonable amount of time ensures the “catching a pokemon” capability is functioning right. This ensures problems are detected and fixed proactively. Synthetic monitoring also helps determine if an issue was due to network latency.
2. Mobile Real User Monitoring
Pokémon Go is a mobile game that is accessed only from mobile devices. Hence Mobile End User Monitoring with crash analytics is imperative to rapidly scope the problem.
Data points – such as how often did crashes occur; what devices, OS and applications versions were being used when the crash occurred; and which geographies did the user come from – are extremely essential to isolate the problem. For example, insights such as “crashes between 6 and 6:30 PM PST were happening from iOS v9 users on West Coast specifically when users attempted to transfer a Pokemon” gives an instant problem scope to delve deeper into.
Further, by tracing individual requests, one can delve into exactly what line of code or what services/microservices could have impacted a particular crash. This data becomes even more insightful if it can be correlated with Twitter sentiment analysis.
A comparison between response time trends and throughput is also another good data point to evaluate if slow responses were due to extra load or an application bug.
3. Server, Database, Application Server Monitoring
In order to deal with scale, the infrastructure to support the game needs to be monitored to spot bottlenecks easily. This requires automatic discovery and health check of all the components that the game runs on.
Considering auto-scaling and high resiliency failover will probably be turned on to cater to the load, the discovery needs to be truly dynamic to track any new nodes that come up. A dynamically discovered topology could have multiple components such as application servers, web servers, databases, load balancers, content distribution networks etc. Memory leaks, CPU consumption, database I/O and space utilization, queues and deadlocks are metrics whose trends need to be monitored continuously with automatic baselines to help identify deviation from normal. Additionally, tracking and correlating log errors via log analysis from these various resources can help diagnose issues rapidly.
4. Predictive anomaly detection for the future
With sudden popularity, one thing that is bound to go out of control is a flood of alerts. To reduce alert noise and ensure that right issues are being worked on, there is the need to have intelligent monitoring alerts. Alerts should be generated based on analyzing, correlating and de-duplicating a set of events and should present sufficient information to enable faster debugging.
As an advanced setup, Pokémon Go monitoring should enable predictive anomaly detection to predict trends on capacity and consumption of backend resources much before they become issues.
Payal Chakravarty is a Program Director of Product Management for IBM Application Performance Management.
