- Quick Highlight
New experimental environment: Tensorflow Serving!
Tensorflow as a widely adopted machine-learning framework and people use it to train their model.
Previously, people have to write a RESTful API server in order to serve requests with trained models.
Tensorflow Serving, in the other hand, is an official tool that features a flexible, high-performance serving system for machine learning models, designed for production environments.
As long as you put trained model at the right directory, the server will load models automatically.
However, you still need to take care how to deploy it to production and setup proper replicas of Tensorflow Serving pods.
Fission now makes these things easier. All you have to do is to create a function with archive of model and a route. Then, Fission will help you to deploy model and scales pods when needed.
Configurable Keep-Alive setting
Previously, due to a known issue Fission disabled Keep-Alive at code-level.
Now, you can enable Keep-Alive by setting
ROUTER_ROUND_TRIP_DISABLE_KEEP_ALIVE to false at router deployment.
Couple things worth noticing:
- This setting increases time for router(s) to switch to newer version for functions that use newdeploy as executor type.
You can prevent this by setting short grace period (
--graceperiod) when creating environment.
- There is an increase in memory consumption of router to keep all active connections.
For details, see PR#1225
Log level through environment variable
All core components now prints Info-Level and above logs by default.
For troubleshooting, you can set env
DEBUG_ENV to true.
For details, see PR#1217
Function updates if config/secret changes
Now, a function will get updated when the referenced configmaps/secrets get updated instead of caching stale data.
Go module support for go environment
Now, go environment supports
go moudle as dependencies management solution.
go module support require fission/go-env-1.12 version and later.#