(I have a vague recollection of reporting a similar task once, but I cannot find it now. If anybody finds it, please merge it into this one.)

At the moment, we log all publish failure events the same way:
https://phabricator.wikimedia.org/diffusion/ECTX/browse/master/modules/eventlogging/ext.cx.eventlogging.js$68

We just save the raw trace JSON data. I have some scripts to parse it, but it's really not a robust way to do it. The field length in EventLogging is limited, and sometimes not the whole trace can be saved, but most of it is not needed anyway. It would be smarter for actual analysis needs, for example T189475, to parse the JSON before logging the event.

Since we are not systematically analyzing the failure data at the moment, and we are rebooting the way we do analysis with the series of tasks that @Pginer-WMF opened recently, we should probably rewrite this part. For T189475, we just need the name of the abuse filter, if it's actually an abuse filter. This can be seen in the trace if it has "code":"abusefilter-disallowed", "code":"abusefilter-warning", or "info":"Hit AbuseFilter. The actual filter name appears in "message":{"key":"NAME"'....