Hi Everyone, I’ve been running into an issue with our experiment setup. I use the API to update experiment and feature flag details, and then link them together. However, on several occasions I’ve noticed that after setting the variations (e.g., x and y), the feature flag later shows both variations as serving “null” (screenshot attached). Could you help me investigate the root cause of this? Many thanks,

Hi there! Nothing comes to mind right away here. Are you able share an example of your experiment/feature api call?

Hi @strong-mouse-55694, thanks for jumping in! Here’s a simplified view of how I’m setting things up through the API: • I create (or fetch) the project and environments. • Ensure SDK connections exist for each environment. • Delete/re-create the feature with a default value (

variationA

). • Create or update an experiment with two variations: ◦

variationA

◦

variationB

• Link the experiment back to the feature by posting a payload with: ◦ a couple of test “force” rules for specific user_ids ◦ then an

experiment-ref

rule, where I map the experiment’s variationIds to the feature values. • Finally, I set the experiment status to

running

. Here’s the relevant snippet for the linking step (the part I suspect may be causing the null issue later on):

{
  "type": "experiment-ref",
  "enabled": true,
  "condition": "",
  "hashAttribute": "user_id",
  "experimentId": experiment["id"],
  "variations": [
    {"variationId": v["variationId"], "value": v["key"]}
    for v in experiment.get("variations", [])
  ]
}

So the feature ends up with rules like: • force →

no-recommendations

• force →

personalize-recommendations

• experiment-ref → falls back to the experiment Do you know if there are any pitfalls with how

experiment-ref

rules need to be structured, or if the mapping of

variationId → value

might be why the flag sometimes reverts to “null” in both variations?

Nothing jumps out, but I'd first check the

variations

array. I'd probably try creating an experiment/flag with hardcoded values as a test case and see if the behavior is the same. If that works, I think there's likely something in how variations is being written.

Thanks a lot @strong-mouse-55694 When I first link the experiment to the feature, the variations look correct for the first couple of days. But after a few weeks, when I check back, some of the features (and not all) end up showing

"null"

being served for both variations.

That's really weird. If you look at the audit log, can you see any changes being made that would change the values?