reducing backlog growth:

As couchup encounters views that take longer than 5 seconds to rebuild, couchup is going to carry on calling additional view urls, triggering their rebuild. Once a number of long running view rebuilds are running even rebuilds that would have been shorter will take at least 5 seconds, leading to a large backlog. If individual databases are large or (map/reduce functions are very inefficient) it would probably be best to set the timeout to something like 5 minutes. If you see more than a couple:

Timeout, view is processing. Moving on.

messages it is probably time to kill couchup and double the timeout.

Observing Index growth

By default view_index_dir is the same as the database directory so if data is in /var/lib/couchdb/shards then /var/lib/couchdb is the configured directory and indexes are stored in /var/lib/coucdh/.shards . You can observe which index shard files are being created and growing or move view_index_dir somewhere separate for easier observation.

What resources are running out?

You can tune couchdb in general, it is hard to say whether tuning is needed once the system is not rebuilding all indexes, etc.
In particular, you would want to look for and disable any autocompaction. Look at files in /proc/[couchdb proc] to figure out the effective fd limits and how many open files there are and whether the crash happens around a specific number of open files. Due to sharding the number of open files is usually a multiple of the number of those in earlier versions.
Look at memory growth and figure out if it is stabalizing enough to use swap to prevent the problem.