This message was deleted.

a hack -- but can you just add a second or two to your activation metric SQL so that the timestamp gets shifted? it won't strictly guarantee the order, but unless you are looking at very short sequences that may be the easiest adjustment

Yes, that is a good hack. Or using the "Advanced Setting" of Metric Delay and setting a negative metric delay, which will operate similarly.

@helpful-application-7107 Hey I’m a colleague auf Paul and we tried this, but hit a limitation which doesn’t make much sense to us: The negative metric delay needs to be <= -0.125 to be recognised in the generated query. If we set value closer to 0 it is simply ignored (eg. -0.05). And even with -0.125 the generated query uses 8min instead of the expected 7.5min 🤔

...
 __experimentUnits AS (
      -- One row per user
      SELECT
        e.userid AS userid
        , (CASE WHEN count(distinct e.variation) > 1 THEN '__multiple__' ELSE max(e.variation) END) AS variation
        , MIN(e.timestamp) AS first_exposure_timestamp
        
        
        , MIN((CASE WHEN a.timestamp >= e.timestamp - INTERVAL '8' minute
      AND a.timestamp <= from_iso8601_timestamp('2024-12-16T12:25:53.665Z') THEN a.timestamp ELSE NULL END)) AS first_activation_timestamp
            
      FROM
        __experimentExposures e
        
        
        LEFT JOIN __activationMetric a ON (a.userid = e.userid)
      
      GROUP BY
        e.userid
    )

Hi, I think we do some rounding under the hood. This isn't totally unexpected to me, but I'll take a look.

Great, thank you Luke!