An LSM like RocksDB has much better write and space efficiency than a B-Tree. That means with RocksDB you will use less SSD than with a B-Tree and the SSD will either last longer or you can use lower endurance SSD. But this efficiency comes at a cost. In the worst-case RocksDB might use 2X more CPU/query than a B-Tree, which means that in the worst case QPS with RocksDB might be half of what it is with a B-Tree. In the examples below I show that RocksDB can use ~2X more comparisons per query compared to a B-Tree and it is nice when results in practice can be explained by theory.But first I want to explain the context where this matters and where it doesn’t matter. It matters when the CPU overhead from RocksDB is a significant fraction of the query response time — so the workload needs to be CPU bound (cached working set). It isn’t a problem for workloads that are IO-bound. Many performance results have been published on my blog and more are coming later this year. With fast storage devices available I recommend that database workloads strive to be IO-bound to avoid using too much memory+power.Basic OperationsWhere does the CPU overhead come from? I started to explain this in my review of the original LSM paper. My focus in this post is on SELECT statements in SQL, but note that writes in SQL also do reads from the storage engine (updates have a where clause, searches are done to find matching rows and appropriate leaf pages, queries might be done to determine whether constraints will be violated, etc).With RocksDB data can be found in the memtable and sorted runs from the L0 to Lmax (Lmax is the max/largest level in the LSM tree. There are 3 basic operations that are used to search for data – get, range seek and range next. Get searches for an exact match. Range seek positions iterators at the start of a range scan. Range next gets the next row from a range scan.First a few comments before I explain the CPU overhead:
I will try to be clear when I refer to the keys for a SQL table (SQL key) and the key as used by RocksDB (RocksDB key). Many SQL indexes can be stored in the same LSM tree (column family) and they are distinguished by prepending the index ID as the first bytes in the RocksDB key. The prepended bytes aren’t visible to the user.
Get is used for exact match on a PK index but not used for exact match on a secondary index because with MyRocks the RocksDB key for a secondary index entry is made unique by appending the PK to it. A SQL SELECT that does an exact match on the secondary key searches for a prefix of the RocksDB key and that requires a range scan.
With RocksDB a bloom filter can be on a prefix of the key or the entire key (prefix vs whole key). A prefix bloom filter can be used for range and point queries. A whole key bloom filter will only be used for point queries on a PK. A whole key bloom filter won’t be used for range queries. Nor will it be used for point queries on a secondary index (because secondary index exact match uses a range scan internally).
Today bloom filters are configured per column family. A good use for column families is to put secondary indexes into separate ones that are configured with prefix bloom filters. Eventually we expect to make this easier in RocksDB.
A point query is evaluated for RocksDB by searching sorted runs in order: first the memtable, then runs in the L0, then the L1 and so on until the Lmax is reached. The search stops as soon as the key is found, whether from a tombstone or a key-value pair. The work done to search each sorted run varies. I use comparisons and bloom filter probes as the proxy for work. I ignore memory system behavior (cache misses) but assume that the skiplist isn’t great in that regard:
memtable – this is a skiplist and I assume the search cost is log2(N) when it has N entries.
L0 – for each sorted run first check the bloom filter and if the key might exist then do binary search on the block index and then do binary search within the data block. The cost is a bloom filter probe and possibly log2(M) comparisons when an L0 sorted run has M entries.
L1 to Ln – each of these levels is range partitioned into many SST files so the first step is to do a binary search to find the SST that might contain the key, then check the bloom filter for that SST and if the key might exist then do binary search on the block index for that SST and then do binary search within the data block. The binary search cost to find the SST gets larger for the larger levels and that cost doesn’t go away when bloom filters are used. For RocksDB with a 1T database, per-level fanout=8, SST size=32M then the number of SSTs per level is 8 in L1, 64 in L2, 512 in L3, 4096 in L4 and 32768 in L5. The number of comparisons before the bloom filter check are 3 for L1, 6 for L2, 9 for L3 and 12 for L4. I assume there is no bloom filter on L5 because it is the max level.
A bloom filter check isn’t free. Fortunately, RocksDB makes the cost less than I expected by limiting the bits that are set for a key, and must be checked on a probe, to one cache line. See the code in AddHash. For now I assume that the cost of a bloom filter probe is equivalent to a few (< 5) comparisons given the cache line optimization.
A Get operation on a B-Tree with 8B rows needs ~33 comparisons. With RocksDB it might need 20 comparisons for the memtable, bloom filter probes for 4 SSTs in L0, 3+6+9+12 SST search comparisons for L1 to L4, bloom filter probes for L1 to L4, and then ~33 comparisons to search the max level (L5). So it is easy to see that the search cost might be double the cost for a B-Tree.The search cost for Get with RocksDB can be reduced by configuring the LSM tree to use fewer sorted runs although we are still figuring out how much that can be reduced. With this example about 1/3 of the comparisons are from the memtable, another 1/3 are from the max level and the remaining 1/3 are from L0 to L4.Range Seek and Next
The cost for a range scan has two components: range seek to initialize the scan and range next to produce each row. For range seek an iterator is positioned within each sorted run in the LSM tree. For range next the merging iterator combines the per-run iterators. For RocksDB the cost of range seek depends on the number of sorted runs while the cost of range next does not (for leveled compaction, assuming uniform distribution).Range seek does binary search per sorted run. This is similar to a point query without a bloom filter. The cost across all sorted runs depends on the number of sorted runs. In the example above where RocksDB has a memtable, 4 SSTs in the L0, L1 to L5 and 8B rows that requires 20 comparisons for the memtable, 4 x 20 comparisons for the L0 SSTs, 21 + 24 + 27 + 30 + 33 comparison for L1 to L5. The total is 235 comparisons. There isn’t a way to avoid this and for short range queries the cost of range seek dominates the cost of range next. While this overhead is significant for short range queries with embedded RocksDB and an in-memory workload it is harder to notice with MyRocks because there is a lot of CPU overhead above MyRocks from optimize, parse and client RPC for a short range query. It is easier to notice the difference in CPU overhead between MyRocks and a B-Tree with longer range scans.
The cost for range next is interesting. LevelDB has a comment that suggests using a heap but the merging iterator code uses N-1 comparisons per row produced which means the overhead is dependent on the number of sorted runs. The cost of range next in RocksDB is much less dependent on the number of sorted runs because it uses an optimized binary heap and the number of comparisons to produce a row depends on the node that produces the winner. Only one comparison is done if the root produces the winner and the root produced the previous winner. Two comparisons are done if the root produces the winner but did not produce the previous winner. More comparisons are needed in other cases. The code is optimized for long runs of winners from one iterator and that is likely with leveled compaction because Lmax is ~10X larger than the next largest level and usually produces most winners. Using a simulation with uniform distribution the expected number of comparisons per row is <= 1.55 regardless of the number of sorted runs for leveled compaction. The optimization in the binary heap used by the merging iterator is limited to remembering the comparison result between the two children of the root node. Were that extended to remembering comparison results between any two sibling nodes in the heap then the expected number of comparisons would be reduced from ~1.55 to ~1.38. For a B-Tree:
The overhead for range seek is similar to a point query — ~33 comparisons when there are 8B rows.
There are no merging iterators. The overhead for range next is low — usually move to the next row in the current leaf page.