Hi GB Team, we are evaluating GrowthBook and while it is the top candidate there are some clarification questions open.
I understood that there are two ways to create an experiment: based on a feature and directly under Analysis/Experiments. Are there special names for these two types of experiments?
In https://docs.growthbook.io/overview#3-experiment-analysis-only 3 use cases are explained. Am I right that experiments based on a features will be used by use case 1 (Full Experimentation Platform) and Experiments without features by use case 3 (Experiment Analysis Only)?
Am I right that neither the SDK endpoint nor the REST API (https://docs.growthbook.io/app/api#authenticated-rest-api-beta) provides information about experiments?
We want to use the experiment life cycle (idea -> draft -> running -> stopped), the analysis support and a lot of other features. The software we want to experiment on is about ML models. So the traffic split / variant assignment will not be done by the frontend or by a growthbook SDK but by a component we are already using: seldon (https://docs.seldon.io/projects/seldon-core/en/latest/index.html). So we are between use case 1 (Full Experimentation Platform) and use case 3 (Experiment Analysis Only). Ingesting the data for analysis with a data source looks feasible. The problematic part is the starting of an experiment:
• the starting of an experiment may fail. I’m not seeing a direct support in the 3 available experiment states.
• the starting of an experiment may take multiple minutes. It would be nice if this could be shown in GrowthBook.
• if the experiment is created without a feature the custom glue code which configures seldon will not get the experiment setup info from an API (see question above). If the experiment is based on a feature one still doesn’t get the experiment setup and the details can differ from the feature.
Can you advise here how to accomplish this?
It would be nice to have a set of dynamic choices for the variants. So the set of trained ML models is queried by GrowthBook and the experimenter can choose from these models in a drop-down box. It looks like this isn’t directly possible. What are “indirect” ways?
12/08/2022, 4:45 PM
There is a data-model. Crazy idea
1. Use kafka connector to get the record that just was created
2. Only update the variants during a idea phase
3. Have the processor query seldon and update the variants directly
1. Requires Kafka connector
2. Writes directly to GB mongo database
3. Eventually you will see the variants
Don;t they have webhooks?
12/08/2022, 6:20 PM
Hello. Our REST api is brand new and we are still adding endpoints to it. Eventually you will be able to create and edit experiment analyses programatically this way. Currently, the only real option is to edit MongoDB directly.
12/09/2022, 11:08 AM
What about using the internal REST API which the frontend uses?
12/09/2022, 5:05 PM
There are probably ways to hack that together, but it will be a little tricky since the authentication system is expecting a browser session and cookies to be available.