In WT-2866 alexander.gorrod says:

However that will just mask another bug: Once the cache is in stuck mode, the cache-full check fires in __wt_cache_eviction_worker, triggering a WT_ROLLBACK for the only operation currently running. There is a problem in the current complex-table code where that WT_ROLLBACK is being swallowed, and the cursor is being left in a state where the primary has key set, but the other column group cursors don't. I think what is happening is that we are retrying the insert, but it's failing because the WT_CURSTD_KEY_SET flag has been cleared.