I have been reviewing latency computation code to try to understand exactly what the max_latency is (still working on that). While reviewing gstbaseaudiosink.c, I came across the following code in the GST_QUERY_LATENCY handling, which seems suspicious:

{{{
          min_latency =
              gst_util_uint64_scale_int (spec->seglatency * spec->segsize,
              GST_SECOND, spec->rate * spec->bytes_per_sample);
          GST_OBJECT_UNLOCK (basesink);

          /* we cannot go lower than the buffer size and the min peer latency */
          min_latency = min_latency + min_l;
          /* the max latency is the max of the peer, we can delay an infinite
           * amount of time. */
          max_latency = min_latency + (max_l == -1 ? 0 : max_l);
}}}

This code seems suspicious to me because in the case where max_l != 0, the final value of max_latency is:
{{{
min_latency_0 + min_l + max_l
}}}
(where min_latency_0 is the value of min_latency returned by gst_util_uint64_scale_int, and min_l and max_l are the latencies for the upstream pipeline)

I am still trying to grasp the exact meaning of max_latency, but it seems very suspicious to me that we would be adding min_l and max_l, so I thought I would flag this computation. 

It seems likely to me that the author intended to be using min_latency_0 in the max_latency computation, and forgot that min_latency had already been operated on.