Yes, you are seeing label switching in your trace. It happens a lot with mixture models.
One of the must-reads related to this topic is Michael Betancourt’s Stan case study Identifying Bayesian Mixture Models. I have a Gist here replicating all the experiment in PyMC3.
As an experiment, you can also switch the trace by hand. We have a discussion here in this post, with code here.
Something related is if you are sampling an Ordered and bounded vector (e.g., x1 < x2 ~ Uniform), current implementation seems to be biased (Sampling uniformly in a triangular support). But for continous RV it should be fine.