Hello and welcome,
This is expected, the output isolator pattern can be a little misleading sometimes, as a blocked output can still block the other inputs.
In your case your first pipeline has an output that sends to 3 other pipelines using the pipeline-to-pipeline communicaton.
One of those pipelines has an Elasticsearch output, so if this output is not working, the logstash memory queue will start to fill up and once it is full it stops accepting new events, this will then propagate to the first pipeline, which will tell the beats input to backoff, this will affect all other outputs in this pipeline.
One alternative would be to use some udp input/output that would drop the evens instead of blocking it, but this also may not be what you want.
This example is described in this similar post.