Metadata unroll operations can fail because of cache pressure in in-memory configurations.
Here's the failure:
[1490812549:504397][5153:0x7f81b57fa700] t WT_SESSION.alter: metadata unroll update file:wt to access_pattern_hint=none allocation_size=512 app_metadata= block_allocation=best block_compressor= cache_resident=false checkpoint= checkpoint_lsn= checksum="uncompressed" collator= columns= dictionary=0 encryption=(keyid= name=) format=btree huffman_key= huffman_value=english id=1 ignore_in_memory_cache_size=false internal_item_max=0 internal_key_max=659 internal_key_truncate=true internal_page_max=32768 key_format=r key_gap=11 leaf_item_max=0 leaf_key_max=818 leaf_page_max=32768 leaf_value_max=0 log=(enabled=true) memory_page_max=5MB os_cache_dirty_max=0 os_cache_max=0 prefix_compression=false prefix_compression_min=4 split_deepen_min_child=0 split_deepen_per_child=0 split_pct=54 value_format=u version=(major=1 minor=1): WT_CACHE_FULL: operation would overflow cache