Interpreting autovacuum logs (9.6)

Hi all,

Please see this graph of data points extracted from autovacuum logs:
https://imgur.com/a/OCRKoDn . It's from a 9.6 instance with default params
for autovacuum with the exceptions of autovacuum_work_mem=10000 and
log_autovacuum_min_duration=100.

1. How should we interpret the # tuples remain reported by the autovac
logs? The source code says it's the "estimated total # of tuples" which to
me means # dead + # live. But that is invalidated by the pattern here where
the orange points (# tuples remain) are dramatically higher than # dead not
removable (blue points) + # dead removed (green points) + # live (which
never exceeded 1M during this entire interval, per count query).

2. Beginning around the first 6/19 tick, what could be causing # tuples
remain to drop steeply after periods of growth when # tuples removed is 0?
I confirmed there was no truncation. And what could the # tuples remain
recurrent asymptote at ~22M mean?

(further context for those curious: the discontinuity at 6/23 is due to an
individual autovacuum run getting stuck. After manually killing that run,
the next one succeeded and you see that reporting toward the right of the
graph)

Thanks!

Gabe