>> crank it up more and delay the checkpoints as much as possible during
>> these updates. 64 segments is already 1024M.
>
> We have 425M rows, total table size is 78GB, so we can imagine a worst
> case UPDATE write is less than 200 bytes * number of rows specified in
> the update (is that logic correct?).
There is also the WAL : all these updates need to be logged, which doubles
the UPDATE write throughput. Perhaps you're WAL-bound (every 16MB segment
needs fsyncing), and tuning of fsync= and wal_buffers, or a faster WAL
disk could help ? (I don't remember your config).
> Inerestingly, the total index size is 148GB, twice that of the table,
> which may be an indication of where the performance bottleneck is.
Index updates can create random I/O (suppose you have a btree on a rather
random column)...