Practical MLOps

Chapter 6. Monitoring and Logging

By Alfredo Deza

Not only is the cerebral anatomy double, and not only is it unarguable that one hemisphere is enough for consciousness; beyond that, two hemispheres following callosotomy have been shown to be conscious simultaneously and independently. As Nagel said of the split-brain, “What the right hemisphere can do on its own is too elaborate, too intentionally directed, and too psychologically intelligible to be regarded merely as a collection of unconscious automatic responses.

Dr. Joseph Bogen

Both logging and monitoring are core pillars of DevOps principles that are crucial to robust ML practices. Useful logging and monitoring are hard to get right, and although you can leverage cloud services that take care of the heavy lifting, it is up to you to decide and come up with a sound strategy that makes sense. Most software engineers tend to prefer writing code and leave behind other tasks like testing, documentation, and very often logging and monitoring.

Don’t be surprised to hear suggestions about automated solutions that can “solve the logging problem.” A solid foundation is possible by thinking thoroughly about the problem at hand so that the information produced is usable. The hard work and solid foundation ideals I describe become crystal clear when you face useless information (it doesn’t help narrate a story) or cryptic (too hard to understand). A perfect example of this situation is a software issue I opened back in 2014 that captured ...