Wednesday, October 13, 2021

Compatible with MySQL or Postgres?

Open and closed source scale-out DBMS that are compatible with MySQL and Postgres have emerged on the market. This is great for the community but there will be much confusion about the meaning of compatible

This post has yet to have anything on the cloud vendors in China. They are doing impressive work but I don't know enough about it. I am happy to update this post when I learn more.

But first, the MySQL and Postgres teams.

One way to describe compatibility is via three levels: protocol, syntax and semantics. By upstream I mean MySQL and PostgreSQL.

  • protocol - an app can use existing client libraries to authenticate and connect 
  • syntax - the DBMS will parse SQL that upstream parses. It is not guaranteed to provide semantics that matches upstream and some syntax can be (or might be) parsed but ignored. Syntax compatible implies a best effort to match upstream semantics but that isn't guaranteed nor must it be guaranteed to be useful. 
  • semantics - the DBMS will match upstream semantics. This implies syntax compatible.
Note that syntax and semantics compatibility aren't all or nothing. A syntax compatible DBMS can be useful without supporting (parsing) 100% of the upstream syntax. A semantics compatible DBMS can be useful without supporting or matching behavior for 100% of upstream syntax.

Also note that semantics compatible implies syntax compatible. But protocol compatible implies neither.

Elsewhere in DBMS land

While this post is about MySQL and PostgreSQL, compatibility is growing in popularity elsewhere:
  • MariaDB provides an Oracle compatible mode that provides syntax but not protocol or semantics compatibility. 
  • EnterpriseDB provides an Oracle compatibility product that I don't know much about. 
  • Amazon will soon open source Babelfish that is protocol compatible with SQL Server.
  • Amazon DocumentDB is protocol, syntax and semantics compatible with (an older version of) MongoDB. It supports some (much) of the MongoDB 4.0 API as of October, 2021 per Wikipedia. Public statements suggest this was built on top of, or reusing, the Aurora PostgreSQL code.
  • added TimescaleDB to Team Postgres
  • added Redshift to Team Postgres
  • added SingleStore to Team MySQL
  • added ClickHouse to Team MySQL
  • added CrateDB to Team Postgres
  • added Dolt to Team MySQL
  • added Team MongoDB
  • added YellowBrick and Greenplum to Team Postgres
  • added Materialize to Team Postgres
  • Building a DBMS to be compatible with MySQL costs more than for Postgres. The Team Postgres projects can reuse BSD licensed PG code while the Team MySQL projects would have to respect the GPL.
  • I have a vague memory of this but to be JDBC compliant the MySQL JDBC driver does a few queries at connect time (either via tables or session/global variables). My JDBC-related bugs are here.

Friday, October 1, 2021

The other way to compress InnoDB: outsource it

There are at least three ways to do compression for InnoDB - classic, holepunch and outsource. 

The classic approach (table compression) was used and enhanced by the FB MySQL team. It might not have been widely used elsewhere. While it works, even for read/write workloads, the implementation is complicated and it isn't clear that it has a bright future.

The holepunch approach (page compression) is much simpler than the classic approach. Alas, I am skeptical that a filesystem will be happy tracking the metadata from doing a holepunch for every (or most) pages written. I am also skeptical that unlink() response times of seconds to minutes (because of the holepunch usage) will be good for a production DBMS. I wrote a few posts about my experience with the holepunch approach: here, here, here and here.

The outsource approach is the most simple from the perspective of InnoDB - let the filesystem or storage do the compression for you. In this case InnoDB continues to do filesystem reads and writes as if pages have a fixed size and the FS/storage compresses prior to writing to storage, decompresses after reading form storage and does all of the magic to make this work. While there are log-structured filesystems in OSS that might make this possible, such filesystems aren't widely used relative to XFS and the EXT family. There is also at least one storage device on the market that supports this.

tl;dr - the outsource approach is useful when the data is sufficiently compressible and the cost of this approach (more write-amp) is greatly reduced when it provides atomic page writes.

After publishing I noticed this interesting Twitter thread on support for atomic writes.

Performance model

I have a simple performance model to understand when the outsource approach will work. As always, the performance model makes assumptions that can be incorrect. Regardless the model is a good start when comparing the space and write amplification with the outsource approach relative to uncompressed InnoDB.


  • The log-structured filesystem has 1+ log segments open for writing. Compressed (variable length) pages are written to the end of an open segment. Such a write can create garbage - the previous version of the page stored elsewhere in a log segment. Garbage collection (GC) copies live data from previously written log segments into open log segments to reclaim space from old log segments that have too much garbage. 
  • The garbage in log segments is a source of space amplification. Garbage collection is a source of write amplification.
  • g - represents the average percentage of garbage (free space) in a previously written log segment. 
  • r - represents the compression rate. With r=0.5 then a 16kb page is compressed to 8kb.
  • Write and space amplification are functions of g:
    • write-amp = 100 / g
    • space-amp = 100 / (100 - g)
Risks in the assumptions:
  • Assumes that g is constant across log segments. A better perf model would allow for variance.
  • Assumes that r is constant across pages. A better perf model might allow for variance.
  • Estimates of write and space amplification might be more complicated than the formula above.

Now I estimate the space-amp and write-amp for outsource relative to uncompressed InnoDB. The ratios are (value for outsource / value for uncompressed InnoDB). For space-amp when the ratio is < 1 then outsource uses less space vs uncompressed InnoDB. For write-amp when the ratio is > 1 then outsource writes more to storage vs uncompressed InnoDB. 

I show below that when the compression rate (r above) is < 0.6 then outsource provides much less space-amp without suffering from a too-large increase in write-amp. But when r is >= 0.6 the increase in write-amp, relative to uncompressed InnoDB, might be a problem.

However, whether a large increase in write-amp is a problem depends on your workload. For example, if 99% of storage IOPs are reads and 1% are writes with uncompressed InnoDB then a write-amp penalty that changes this from 1% writes to 2% writes is unlikely to be a problem.

I created a graph with Desmos to show the three ratios. The graph allows r to be adjusted. I then copied some values from the graph into a table below. The graph has 3 curves, one for each ratio:
  • s is the space-amp ratio where s = r * 100 / (100 - g).
  • w1 is the write-amp ratio assuming the doublewrite buffer is enabled where w1 = r * 100/g.
  • w2 is the write-amp ratio assuming the doublewrite buffer is disabled. This assumes that outsource provides atomic page writes for free. The formula is w2 = r/2 * 100/g.
The table below has values from the graph for r = 0.4, 0.5 and 0.6. What I see in the graph is that with r in (0.4, 0.5) and the doublewrite buffer disabled (w2) it is possible to get much of the compression benefit (see s) from outsource without a significant increase in write-amp. But the write-amp penalty can be a problem when r >= 0.6. Of course, whether or not more write-amp is an issue depends on the storage read and write rates as I explained above. 

g       s       w1      w2
20      0.50    2       1
30      0.57    1.33    0.67
40      0.67    1       0.5

g       s       w1      w2
20      0.63    2.5     1.25
30      0.71    1.67    0.83
40      0.83    1.25    0.63

g       s       w1      w2
20      0.75    3       1.5
30      0.86    2       1
40      1       1.5     0.75

Wednesday, September 29, 2021

Automating memory management in a DBMS

I recently read two papers on automating memory management - one for DB2, one for Oracle. While the papers aren't new (~2005), they are definitely worth reading. AFAIK both made it into the products, I have no idea how that turned out but the outcome doesn't change my opinion that the papers were excellent. Disclaimer - I worked on query execution for Oracle at the time and made a few changes to the row sources that I maintained to support the work described by the Oracle paper.

Both describe methods to automate memory management with the goal of improving performance. In many DBMS memory management is difficult for several reasons:

  • It isn't global and the DBA must estimate good values for the sizes of various caches and other large memory consumers: sort and hash for order by, aggregation, join and index create.
  • It isn't bounded when allocations are specified per instance of sort or join rather than one limit on the memory used by all sorts or joins. Too many concurrent queries == OOM in this case.
  • It is static. This is an issue for large caches like the buffer pool (block cache).
There was a marketing battle between Oracle and IBM in the mid-2000s over autonomic computing and these papers, along with getting them into products, is one outcome from that battle.

From the DB2 paper

A summary:
  1. Differentiable functions were created to explain the query latency saved per page of memory for a variety of consumers. All functions had the same form (see section 3.2) with graphs that look like this. I think that time saved means wall clock time to account for CPU and IO overheads.
  2. For a few caches (statement and buffer pool) online simulators were added to the DBMS to estimate changes to the hit rate if the cache were given more memory. 
  3. Constrained optimization was used to determine the optimal allocation at any point in time. The constraint was the amount of memory available. I assume that Lagrange multipliers were used.
  4. Feedback control was used to apply the desired memory configuration. One benefit from this is to avoid negative impacts from suddenly applying significant changes.
  5. Decisions are revisited because workloads (types of queries, concurrency) changes.
  • At least for sort, the real curve for time vs memory can't be described by a differentiable function. See page 2 in the Oracle paper: there are three intervals where each interval can be explained by a function but the points where the intervals meet are a problem. I am not sure whether constrained optimization can handle that.
  • At a higher level, the paper doesn't consider the steps in the time vs memory benefits for some row sources (this is a kind of a repeat of the previous comment).
  • It wasn't clear whether row sources could give back memory. There is a risk from giving a long-running sort or hash a lot of memory. If there is no facility to get back some of that memory then memory allocation will be far from optimal until that query completes or is killed.
From the Oracle paper

While Oracle has automatic memory management for caches (plan cache, buffer pool) and queries (hash, sort, bitmap index operators) the paper is limited to memory management for queries (PGA in Oracle terminology).

  1. for each instance of a row source (a query is composed of multiple row sources, some row sources can use a lot of memory for sort and hash) the knees in the response time vs memory graph are estimated, see section 3.2 in the paper. These knees are a function of the row source and the data specific to a query (query A might be able to sort in-memory with 100M of RAM while query B might require 1G to remain in memory).
  2. decisions are made about the amount of memory that each row source can used based on the information from the previous point
  3. over time a row source might be able to get more memory and might be told to give back memory. Giving back memory might not be immediate, but should eventually be done.
  • the paper was vague about how point #2 was done. While section 4.2.2 lists 5 rules used to guide the decisions it wasn't clear there was a goal beyond making sure all queries have enough memory to run. In contrast the IBM paper was clear that constrained optimization was used and explains the function that was optimized.

Tuesday, September 28, 2021

Developer experience, what about the other stakeholders?

While developer experience gets a lot of press, there are four stakeholders when you provide a database service. So we can call these DX, UX, MX and OX for developer, user, management and operations experience:

  • developers want the DBMS to stay out of their way. Schema is one example because waiting for a schema change gets in the way. Note that NoSQL databases have less schema rather than being schema-less because indexes are schema. I am curious whether less schema leads to a great developer experience in the long run for large scale projects given the risk of an unmanaged schema and poorly understood data. The developers who were able to move fast early in the project can create much tech debt for those who arrive years later.
  • users of the services that depend on the database want great QoS - high uptime, low and predictable latency for queries, low chance of lost data. They just want to use your database-backed app without problems.
  • management wants to minimize cost while getting great QoS.
  • operations wants to be able to sleep when they are oncall (self-healing database, auto failover, etc). It helps if the database isn't in chaos mode during working hours as that gives them freedom to get their work done.

Tuesday, September 21, 2021

Review of DiffKV

This is a review of Differentiated Key-Value Storage Management for Balanced I/O Performance that was published in ATC 2021. This is a wonderful paper that resolves several of my concerns about key-value separation (here and here). The authors have experience with key-value separation via Titan which is part of TiDB.

The key idea in the paper is to use leveled compaction for keys and tiered compaction for values. Write-amplification is reduced by not using leveled compaction for values. Read-amplification is reduced, relative to classic key-value separation (see WiscKey), by using tiered compaction for values rather than logs. With classic key-value separation the worst case read-amp for a scan is the need to do a random read from the log for every qualifying key as that random read can include a block decompression and a storage IO. With DiffKV the use of tiered compaction for values provides some ordering to avoid that worst case.

While I enjoyed the paper, it didn't have performance results with compression enabled and I am curious about the impact of decompression overhead on point and range read latency.

A summary of the approach:

  • small values are stored inline in the LSM tree
  • large values are stored in logs, called vLogs. The paper explains a few improvements to make GC faster in this case. But mostly this is classic key-value separation.
  • medium values use the vTree (tiered compaction for values)
  • the user defines the two size limits that determines whether a value is small, medium or large
The vTree

Medium values are stored in the vTree which resembles an LSM that uses tiered compaction. A vTable is the unit of storage in the vTree. A vTable is 8MB by default and stores KV pairs in key order. A sorted group is a sequence vTables for which the keys don't overlap -- (vTable, sorted group) are similar to an (SST, sorted run) in RocksDB. A vTree has multiple levels where each level has one or more sorted groups and the levels have exponentially increasing size.

This is similar to tiered compaction for two reasons. First, there are multiple sorted groups per level. Second, when merges are done to move values from level N to N+1, the merge process only reads data from level N and then writes it to level N+1. It doesn't read and rewrite data already on level N+1. 

To reduce the number of times that values get rewritten the vTree has fewer levels than the LSM tree that stores keys. The smallest N levels in the key's LSM tree share the smallest vTree level. I assume the remaining levels in the key's LSM tree each have their own level in the vTree.

vTree GC

Keys in the LSM tree point into the vTree. When a value is moved between levels in the vTree the key's entry in the LSM tree must be updated to point to the new location in the vTree. DiffKV couples GC in the vTree with GC in the key's LSM tree to avoid extra writes. By coupling I mean that vTree GC is extra work added to compaction done on the key's LSM tree. When a key is moved from level N to level N+1 in the LSM tree then its value might be moved to the next level in the vTree.

By coupling it like this, DiffKV avoids the need to probe the index (key's LSM tree) to determine whether a value is live -- which is an extra overhead that occurs with classic key-value separation. It also avoids generating extra writes to change the point into the vTree that is stored with the key.

This above is called compaction-triggered merge and driven by compaction of the key's LSM tree. Another reason for doing vTree GC is called scan-triggered and triggered by scans that encounter regions of the vTree that need GC.

I am curious whether there is a need to also trigger GC based on vTables that have excessive space-amplification.

vTree rewrites

Leveled compaction has more write-amp than tiered because it rewrites previously written KV pairs. While vTree GC frequently avoids the need to do rewrites, there is one case where a rewrite is needed and I didn't learn enough about how DiffKV handles this.

There can be space wasted from vTables that have low utilization rates because most of the values in the vTable have been deleted. In this case something should be done to copy out (rewrite) the values to a new vTable. And when values are moved the key LSM tree must be updated to reference their new location. For workloads with updates and deletes this will be needed in the largest level of the vTree and might be needed for smaller levels.

Saturday, August 14, 2021

Happy 15th to Percona

I am thrilled to be celebrate Percona's 15th anniversary. My time with MySQL began about the same time as the founding of Percona. Those years, the mid-2000s, were the dark ages for MySQL. There was doubt about the future of MySQL because there were many things that needed to be made better. Fortunately, upstream and the community, with much help from Percona, rallied to fix the problems and add needed features.

Percona has been a huge part of the community. I don't have time to list everything, but examples include saving the user conference, providing an open-source hot backup solution, educating many of us (including me) via their blog posts and helping push the product to get so much better. I have also been a customer as they helped with the MyRocks effort and in the great patch migration (5.6 to 8.0).

I like that the anniversary book starts by mentioning the multi-core scaling problem that InnoDB had in the mid-2000s. My MySQL deployment used 4-core, 1-socket CPUs at the time and 8-core, 2-socket servers were about to arrive. It was difficult to get more than 10k QPS from MySQL/InnoDB on such hardware regardless of the number of cores or sockets. InnoDB didn't benefit from 2-socket servers because of contention in the custom InnoDB mutex (which was also used to guard state in the custom InnoDB rw-lock). A fix was first proposed by Yasufumi Kinoshita, who would soon join Percona. After seeing his presentation at the user conference, my team at Google proposed a similar solution which was accepted by the InnoDB team and a serious problem was resolved.

I also like that the book is full of stories from people in the community. I know many of these people from my time traveling to conferences. I am an introvert except when at conferences. I rarely observe talks because I am out in the hallway track talking to others.

Wednesday, August 11, 2021

On storage engines

I wish it were easier to implement new storage engines for MySQL, Postgres and MongoDB and other OSS databases. There is so much innovation that we miss out on - FASTER is one example. All (MySQL, MongoDB, Postgres) have storage engine APIs but there are not many OSS implementations of them.

MyRocks, MySQL Aurora and MySQL HeatWave are examples of the benefits. But they also show that it helps to have the backing of a well-funded company because this is a huge undertaking.

The API for Postgres is the most recent and perhaps that will be the most popular. The API for MySQL is the oldest and has become harder as more requirements (like partitioning) are pushed into it. The MongoDB API is friendlier than the MySQL API, but MongoRocks was deprecated when the API was enhanced to support transactions and RocksDB wasn't able to support user-provided commit timestamps.

One problem is naming. MongoDB and WiredTiger are databases, but WiredTiger is also a component of MongoDB. To avoid confusion I will use storage engine for things like WiredTiger, RocksDB and InnoDB and database for the systems that provide query languages, replication and more.

I have been involved in three such engines -- MyRocks, MongoRocks and a read-only MySQL engine used long-ago at Google. MyRocks benefited from a large team and Sergey Petrunya at MariaDB and MongoRocks benefited from amazing support from MongoDB the company.

Two interesting things are:

  • The impact of database design decisions on the storage engine API. See what is needed for transactions in the MongoDB API.
  • The impact of the original storage engine on the storage engine API.
    • MongoDB doesn't take advantage of clustered indexes in WiredTiger or RocksDB. An extra (hidden) index must be used to map between DiskLoc and the PK index. See SERVER-14569.
    • I have more research to do before I understand the impact of vacuum and 32-bit transactions IDs on the Postgres API.

I also wonder whether the RocksDB API could be the universal storage engine API. In theory any storage engine implementing that would be able to replace RocksDB in MyRocks or MongoRocks. The goal is then to implement the RocksDB API glue once per database and then be able to use a variety of storage engines based on that. Of course, XKCD has a great take on standards and this might just be naive and wishful thinking on my part.

Saturday, March 13, 2021

Sequential IO and an LSM tree

I see statements that sequential IO is a benefit of using an LSM tree. That is a vague and truthy statement. It is correct that the writes done by compaction are sequential per-file. But I am interested in the IO from the perspective of the storage device. 

The truthier claim is that with an LSM there will be many streams of IO (read & write) that benefit from large IO requests. The reads and writes done by compaction are sequential per file, but there is much concurrency and the storage device will see many concurrent streams of IO.

The IO patterns for an LSM with a busy workload are:

  • N compaction threads where each thread is reading from ~10 files, merging the results and writing the output to create new SSTs. Each thread has one file open at a time that is written sequentially. The reads are sequential per-file. At any point in time there can be 10*N read streams and N write streams. The reads benefit from large IO requests and prefetch. The writes benefit from async write-back. The writes might benefit from clever placement on the storage device (see multistream). The writes are likely to generate large requests to the storage device.
  • The WAL, if enabled, is another write stream. If fsync is done on commit then stream gets a sequence of small writes.
  • User queries generate read requests. For OLTP these are mostly small (random) requests. But in some cases (logical backup, analytics) there can be scans.

Sunday, March 7, 2021

How to submit a PR for RocksDB

I submitted my first PR to RocksDB as an external contributor and have advice on the process.

  • Read the Contribution Guide
  • Run make format on your diffs. This depends on clang-format and you might need to install it.
  • Edit if there are user-facing changes
  • Optionally run make check to confirm tests pass. Some tests failed because my small servers have 16G of RAM, the tests use /dev/shm and needed more than the 8G that /dev/shm gets on my servers.
  • Confirm that RocksDB LITE builds still work: LITE=1 make db_bench
  • Some of the CI tests are flaky now and have bogus failures (see here). To get CI tests to repeat I edit and push small changes to The team is fixing the intermittent test failures. Alas, one of the intermittent problems was from internal gcc failures on ARM -- something I can't reproduce on my x86 hardware. The good news is that after three attempts everything passes.
  • Google for advice on how to do this in git. I am not a git guru but the advice I find is great

Friday, February 26, 2021

More on read-only benchmarks and an LSM

This adds to my last post on the impact of LSM tree shape on read performance. I updated the test script to add one more stage, l0_and_lmax, that runs after compact_all. A full compaction is done for compact_all to show the benefit (more QPS) of reducing the number of trees to search to 1. For the l0_and_lmax stage overwrites are done to get 6 SSTs in the L0. At this point there is data in the L0 and max level of the LSM tree -- thus the name l0_and_lmax.

The updated results and changes to my test scripts are here. The spreadsheet with data and charts is here.

The summary is that query performance suffers if you do all of the work to fully compact the LSM tree and then put 6 SSTs in the L0. By suffers, I mean it is almost as bad as when the LSM tree has data everywhere (memtable, L0, all levels after the L0). This is visible in the charts below -- compare l0_and_lmax with pre_flush and post_flush.


Wednesday, February 24, 2021

Read-only benchmarks with an LSM are complicated

I often see benchmark results for an LSM tree where the shape of the tree is not described. When the results are worse than desired I wonder how much of that is inherent to the LSM tree versus the configuration. It might be easier to make an LSM tree look bad than to make a B-Tree look bad.

I have results from RocksDB for which QPS varies by up to 5X depending on the shape of the LSM tree. The workload is read-only and CPU-bound. The impact of the LSM tree shape on CPU overhead is known but not widely known. Hopefully this helps with that.

Read amplification for an LSM is usually worse than for a B-Tree. But read-amp should be described separately for IO vs CPU because (usually) CPU read-amp is the problem. The unit for IO read-amp is pages read from storage (real storage, not a read from the OS page cache). The unit for CPU read-amp is CPU overhead but some combination of cache lines read, keys compared and bloom filters checked per query might be a good proxy.

The extra CPU read-amp comes from the number of trees in the LSM tree that must be searched on queries. Perhaps LSM tree is a misnomer because it usually has more than one tree, but perhaps not enough trees for LSM forest to be a good name. Is it too late to rename LSM tree to LSM trees? The memtable is one tree, each sorted run in the L0 is another tree, and then each level beyond L0 is one more tree. For the configurations I use it is common for there to be from 10 to 20 trees in one RocksDB instance. While bloom filters help, they don't eliminate the overhead. Even with a bloom filter some searching is done (cost is a function of |values|) and then there is the overhead of the bloom filter check. For too much information on the CPU overhead of a RocksDB query, see this post from me.

I will begin using LSM tree shape to describe the state of the LSM tree where the state includes:

  • number of KV pairs in the active memtable
  • number of immutable memtables waiting to be flushed
  • number of sorted runs in the L0
  • number of levels beyond the L0, and their sizes
Just keep in mind that it is hard to interpret performance results for a CPU-bound and read-only workload without understanding the LSM tree shape. It is easy to make the LSM look better or worse -- on purpose or by accident.

Finally, LSM implementations can get better at this by adapting their LSM tree shape to become more read-friendly when a workload shifts to read-only or read-mostly phases.

Performance summary

The tests I do change the per-query CPU overhead. I directly measure and report throughput but it, per-query response times and per-query CPU overhead are correlated.
  • Flushing the memtable has no impact on throughput
  • The impact from too many trees to search is worse for workloads that don't use bloom filters (like range queries, or any queries when bloom filters don't exist). The worst-case was ~5X for range and point without bloom vs ~2X for point with bloom.
  • There are incremental benefits from compacting the L0, compacting the L1 and doing full compaction. Each of these reduces the number of trees to search.
  • Bloom filters in an LSM aren't O(1). There is a search cost, O(logN), to find per-SST filter
  • Feature request for LSM trees - adapt to be more read-friendly when a workload is read-heavy
  • This is more of an issue for read-only benchmarks because the LSM tree shape is usually stuck in the same state for the duration of the test. That state might be good or bad and this can add noise or misleading results to your tests. For a test with writes the LSM tree is likely to have data in all places (memtable, L0, L1, ...) most of the time which might not help read performance but will reduce read response time variance.

A spreadsheet with absolute and relative throughput is here. All results are from the configuration with the per-level fanout set to 8.

My tests

I ran tests for three types of queries: short range queries, point queries with bloom filters and point queries without bloom filters. I repeated the queries for four LSM tree shapes described below. By data is in the memtable I mean that the memtable is at least half full but there were no immutable memtables. By data is in L0 I mean there were six SSTs (sorted runs) in L0. 

  • pre_flush - data is in the memtable, L0 and each level beyond L0
  • post_flush - memtable is empty while L0 and larger levels have data
  • compact_l0 - memtable and L0 are empty while L1 and larger levels have data
  • compact_l1 - memtable, L0 and L1 are empty while L2 and larger levels have data
  • compact_all - all data is in one sorted run on Lmax (the max level of the LSM tree)
The query tests are provided by db_bench. I used seekrandom for range queries with --seek_nexts=8 to fetch 8 values per query. I used readrandom for point queries with and without bloom filters. Elsewhere I call these tests range query, point with bloom and point without bloom.

Tests were repeated after loading 1M, 10M, 100M and 200M 128-byte values. All tests used one thread and were run on my small servers. The servers have 16G of RAM and RocksDB used an 8G block cache. The database fit in the block cache for all tests except for 200M values. Test output, command lines, and a diff to db_bench are here. It wasn't easy to get db_bench to cooperate. Some of the parts of the db_bench diff will be useful for future tests and I hope to improve and submit a PR for it. The useful features include new benchmarks (things that go in --benchmarks=...) including:
  • flush - flushes the memtable
  • compact0 - compacts L0 into L1 (and blocks until done)
  • compact1 - compacts L1 into L2 (and blocks until done)
  • waitforcompaction - a hacky attempt to block until compaction is done. Can be used as --benchmarks=fillrandom,compactall,waitforcompaction,readrandom because compactall doesn't block until compaction is done.
Without the options above it was hard to get the LSM tree shape into a deterministic state. Once I had the new features it took a few experiments to find the correct number of overwrites to do to make both the memtable and L0 almost full (275000), while also avoiding doing too many flushes and triggering an L0->L1 compaction.

One interesting problem was that immediately after fillrandom the number of sorted runs in the L0 was almost always zero regardless of the number of values written. The reason is that I run tests with fsync on commit disabled and writes were so fast that at the end of fillrandom there were always too many sorted runs in L0. Once fillrandom ended a compaction was scheduled and it merged all L0 sorted runs into L1. The workaround is to run db_bench twice -- once to do fillrandom, flush memtables and then compact L0. Then again to run overwrite to put some SSTs in L0 and then do query tests. This is in the script.

# Step 1
db_bench ... --benchmarks=fillrandom,flush,waitforcompaction,compact0

# Step 2
db_bench ... --use_existing_db=1 --benchmarks=overwrite,waitforcompaction,...

LSM tree shape

Below I describe the LSM tree shape using logical rather than physical names for the levels because an artifact of dynamic leveled compaction is that compaction stats might show L0 followed by L4 where L4 is logically L1.  

The results show something that I hope can be improved. When configuring RocksDB you can reason about the read, write and space amplification tradeoffs but then the following happens as the LSM tree grows and your assumptions might be far from the truth. The problem is visible in the 10M and 100M values tests when the per-level fanout is 10 (the default). They use the same number of levels while the 100M test has 10X more values.

For all tests the target size for L1 is 64M and dynamic leveled compaction was enabled. But the interaction between dynamic leveled and my configuration means that the LSM trees for the 10M and 100M values tests use the same number of levels with a dramatically different sizeof(L1) / sizeof(L0) ratio. That ratio is <= 2 for the 10M values test vs ~10 for the 100M values test. The LSM tree uses one too many levels in the 10M values test. 

LSM tree shape by number of values inserted when per-level fanout is 10 (default):
  • 1M
    • memtable is ~80% full with 38,239 values
    • L0 has 6 sorted runs and 33M of data
    • L1 has 87M of data
  • 10M
    • memtable is ~80% full with 38,199 values
    • L0 has 6 sorted run and 34M of data
    • L1 has 80M of data
    • L2 has 829M of data
  • 100M
    • memtable is ~80% full with 38,240 values
    • L0 has 6 sorted runs and 39M of data
    • L1 has 792M of data
    • L2 has 7.77G of data
  • 200M
    • memtable is ~80% full with 38,215 values
    • L0 has 6 sorted runs and 38M of data
    • L1 has 163M of data
    • L2 has 1.61G of data
    • L3 has 16.12G of data
I then repeated the test with the per-level fanout reduced to 8 via the max_bytes_for_level_multiplier option in db_bench. In this case the number of levels is what I expect.
  • 1M
    • memtable is ~80% full with 38,209 values
    • L0 has 6 sorted runs and 33M of data
    • L1 has 87M of data
  • 10M
    • memtable is ~80% full with 38,241 values
    • L0 has 6 sorted runs and 34M of data
    • L1 has 101M of data
    • L2 has 820M of data
  • 100M
    • memtable is ~80% full with 38,264 values
    • L0 has 6 sorted runs and 39M of data
    • L1 has 126M of data
    • L2 has 1.0G of data
    • L3 has 7.9G of data
  • 200M
    • memtable is ~80% full with 38,215 values
    • L0 has 6 sorted runs with 38M of data
    • L1 has 251M of data
    • L2 has 1.98G of data
    • L3 has 15.83G of data

Thursday, February 18, 2021

Still not finding CPU regressions in RocksDB

I repeated the search for CPU regressions in recent RocksDB releases using a m5ad.12xlarge server from AWS. I was happy to not find problems, just as I did not find any on a small servers.

I setup the AWS host using this script and then ran the script for 1, 4, 8 and 16 concurrent clients. I forgot to save the command line, but the only thing I changed was the number of background threads that RocksDB can use (increased from 2 to 8).

The AWS host has two direct attached NVMe drives and I used one for the database directory with XFS and discard enabled. I also disabled hyperthreads when I setup the host.

Results are in github for one, four, eight and sixteen threads. The output is created by the script and has 3 sections. The first is the response time/operation per test. The second is the throughput per test. The third is the relative throughput per test (relative to RocksDB version 6.4). The versions tested are (v64=6.4, v611=6.11, ..., v617=6.1.7).

Below I only show the output from the third section. A value > 1 means the new version has more throughput than RocksDB version 6.4. The values tend to be close to 1 and I don't see significant regressions.

My focus is on the first two and last 5 tests and the others are read-only and can suffer from noise because the state of the LSM tree isn't deterministic (data in memtable and L0 can vary). In the long run I hope for options in db_bench to flush the memtable and compact L0 into L1 to reduce this problem.

Results for 1 thread

v64 v611 v612 v613 v614 v615 v616 v617 test
1 0.99 0.99 0.99 0.99 1.00 0.98 1.00 fillrandom
1 0.97 0.96 0.97 0.96 0.97 0.97 0.97 overwrite
1 0.89 0.94 0.95 1.01 0.94 0.90 0.91 readseq
1 0.79 0.77 0.82 0.80 0.81 0.69 0.76 readrandom
1 0.98 1.00 0.99 0.98 1.02 0.67 0.99 seekrandom
1 0.89 0.98 0.96 0.95 0.96 0.89 0.94 readseq
1 0.85 0.79 0.80 0.79 0.81 0.80 0.82 readrandom
1 0.92 0.97 0.96 0.96 1.00 1.09 0.98 seekrandom
1 0.96 0.98 0.96 0.91 0.97 1.10 0.97 seekrandom
1 0.96 0.96 0.96 0.97 1.00 1.07 0.96 seekrandom
1 1.01 1.04 1.05 1.00 1.07 1.09 0.99 seekrandom
1 1.06 1.01 0.87 0.98 0.95 1.03 1.00 readwhilewriting
1 0.84 0.94 0.98 1.03 0.96 0.94 0.97 seekrandomwhilewriting
1 0.97 1.03 0.97 0.98 0.99 0.97 0.96 seekrandomwhilewriting
1 0.93 0.97 0.93 0.97 0.94 0.97 0.99 seekrandomwhilewriting
1 0.95 0.97 1.02 0.95 1.01 1.01 1.01 seekrandomwhilewriting

Results for 4 threads

v64 v611 v612 v613 v614 v615 v616 v617 test
1 0.99 0.98 0.99 0.98 0.99 0.98 0.98 fillrandom
1 0.99 1.07 1.09 1.03 1.06 1.02 1.02 overwrite
1 0.92 0.94 0.94 0.97 0.96 0.95 0.93 readseq
1 0.84 0.90 0.90 0.90 0.74 0.74 0.99 readrandom
1 0.72 1.01 0.98 1.01 0.72 0.84 1.04 seekrandom
1 0.90 0.97 0.90 0.89 0.88 0.90 0.90 readseq
1 0.87 0.94 0.94 0.95 1.03 0.86 0.95 readrandom
1 0.75 1.06 0.99 1.01 1.11 0.86 1.03 seekrandom
1 0.73 1.02 0.97 0.97 1.07 0.79 0.95 seekrandom
1 0.76 1.02 0.96 1.02 1.12 0.85 1.03 seekrandom
1 0.78 0.95 0.96 0.96 1.05 0.82 0.96 seekrandom
1 0.94 0.93 0.91 0.94 0.95 0.94 0.96 readwhilewriting
1 0.96 0.92 0.94 0.94 0.94 0.92 0.97 seekrandomwhilewriting
1 0.98 0.94 0.98 0.96 0.98 0.97 0.97 seekrandomwhilewriting
1 0.97 0.93 1.04 0.98 0.98 0.95 0.96 seekrandomwhilewriting
1 0.96 1.02 1.00 1.02 0.98 1.01 0.98 seekrandomwhilewriting

Results for 8 threads

v64 v611 v612 v613 v614 v615 v616 v617 test
1 1.01 0.96 0.97 0.99 0.97 0.97 0.99 fillrandom
1 0.74 0.84 1.07 1.01 1.00 0.99 0.99 overwrite
1 0.89 0.89 1.09 1.06 1.10 1.08 1.08 readseq
1 1.01 1.01 0.89 0.97 1.01 0.72 0.88 readrandom
1 1.17 0.99 1.02 0.94 0.98 0.80 0.94 seekrandom
1 1.01 1.06 1.01 1.01 0.99 0.98 0.97 readseq
1 0.98 0.99 0.91 0.96 0.92 0.81 0.92 readrandom
1 1.13 0.92 0.95 0.88 0.92 0.74 0.89 seekrandom
1 1.12 0.94 0.96 0.91 0.96 0.77 0.91 seekrandom
1 1.15 0.96 1.00 0.94 0.95 0.79 0.92 seekrandom
1 1.04 0.96 0.91 0.92 0.92 0.81 0.90 seekrandom
1 0.97 0.97 0.99 0.99 0.98 1.02 1.00 readwhilewriting
1 0.98 1.01 0.98 0.96 0.97 0.97 0.98 seekrandomwhilewriting
1 0.95 0.94 0.95 0.96 0.95 0.92 0.94 seekrandomwhilewriting
1 0.94 0.97 0.96 0.96 0.95 0.95 0.98 seekrandomwhilewriting
1 0.98 1.01 0.98 0.96 0.96 0.94 0.92 seekrandomwhilewriting

Results for 16 threads

v64 v611 v612 v613 v614 v615 v616 v617 test
1 0.99 0.98 0.98 0.95 0.97 0.96 0.99 fillrandom
1 0.93 0.95 1.41 1.29 1.32 1.30 1.32 overwrite
1 0.95 0.99 1.12 1.13 1.12 1.09 1.12 readseq
1 1.07 0.85 0.97 0.95 0.88 0.95 0.91 readrandom
1 1.03 1.00 1.00 1.03 0.96 1.02 1.02 seekrandom
1 0.98 1.06 0.97 0.97 0.96 0.97 0.94 readseq
1 1.04 0.88 1.03 1.01 0.97 0.96 1.05 readrandom
1 0.99 1.00 0.98 0.99 0.94 0.97 0.99 seekrandom
1 1.00 1.00 0.99 1.00 0.93 0.98 1.00 seekrandom
1 1.00 0.99 0.97 1.01 0.93 0.97 0.98 seekrandom
1 0.98 0.98 0.98 0.97 0.93 0.94 0.95 seekrandom
1 0.93 0.97 0.97 0.97 0.94 0.96 0.93 readwhilewriting
1 0.97 0.97 0.92 0.96 0.99 0.97 0.98 seekrandomwhilewriting
1 0.95 0.97 0.96 0.97 0.95 0.98 0.94 seekrandomwhilewriting
1 0.97 1.00 0.99 0.97 0.99 0.99 0.94 seekrandomwhilewriting
1 0.97 1.01 1.00 1.02 1.00 0.98 0.99 seekrandomwhilewriting

Monday, February 15, 2021

Not finding CPU regressions across RocksDB releases

I looked for CPU regressions in recent RocksDB releases and was happy to not find them. 

My workload was low concurrency (1 or 2 threads), an in-memory workload and a small server. I tested RocksDB versions 6.4 and 6.11 through 6.17. My servers use Ubuntu 20.04 which has g++ version 9.3. I wasn't able to compile versions prior to 6.4 because there were compiler errors that I didn't try to resolve. Such is the cost of using modern C++.

Running the tests

I used the, and scripts from my github repo. The scripts do two things for me. First, they handles changes to the db_bench options across RocksDB versions. Second, runs tests in a sequence that is interesting to me. I need to update for changes in the 6.X branch with respect to the db_bench options.

I forgot to do ulimit -n 50000 prior to running tests and repeated tests after fixing that. I have forgotten to do that many times in the past.

I ran the test in two modes: not-cached and cached. By not-cached I mean that the database is larger than the RocksDB block cache and cached means the database fits in the RocksDB block cache. In both cases all data is in the OS page cache.

For the benchmarks:10M KV pairs were inserted in the initial load, each test step was run for 300 seconds and the LSM tree was made small (8M write buffer, 32M L1) to get more levels in the LSM tree.

Command lines that are useful to me, but the scripts might be inscrutable to you:

# To run for cached
bash 100000000 64 300 8 32 $(( 10 * 1024 * 1024 ))
# To run for not-cached
bash 100000000 64 300 8 32 $(( 1 * 1024 * 1024 ))
# To generate summaries for response time and throughput
bash v64 v611 v612 v613 v614 v615 v616 v617


While there are many test steps (test step == one run of db_bench) the most interesting are the first two (fillrandom, overwrite) and the last 5 (readwhilewriting, seekrandomwhilewriting with different range sizes). Results can be misleading for the read-only tests that are done in between these as the performance on them depends on the shape of the LSM tree. The amount of data in the memtable, L0 and L1 isn't deterministic and can have a large impact on the CPU overhead for queries. In MyRocks I reduce the impact from this by flushing the memtable and compacting the L0 but db_bench doesn't have options for that (yet). It does have an option to do a full compaction but that is too much for me.

So I will share the results for all test steps but focus on the first 2 and last 5. There aren't significant regressions from v64 to v617. Results with a larger font and numbers for both response time and throughput are in github for cached and not-cached.

This has the QPS from the cached test:

v64     v611    v612    v613    v614    v615    v616    v617    test

357763  339811  329584  342793  339205  347598  339990  348218  fillrandom
344379  344567  340444  330834  325129  333669  331436  347029  overwrite
2817650 2779196 2877354 2886832 2887711 2774030 2710212 2823124 readseq
159063  129527  121106  123190  121325  134453  120146  137377  readrandom
77047   73334   64849   71194   49592   93552   64414   98480   seekrandom
3760630 3319862 3436593 3424777 3416794 3468542 3348936 3419843 readseq
216189  168233  170755  167659  177929  194189  170752  197241  readrandom
77176   74307   67099   73279   83014   95671   66752   97165   seekrandom
76688   73207   65771   71620   80992   93068   64924   94364   seekrandom
67994   65093   59306   65372   72698   80427   59296   83996   seekrandom
35619   33270   32388   34049   36375   38317   32091   38652   seekrandom
155204  151360  150730  151218  149980  150653  150261  148748  readwhilewriting
57080   54777   56317   55931   55271   55564   55581   56334   seekrandomwhilewriting
56184   53540   54838   54445   54450   54633   54465   54410   seekrandomwhilewriting
51143   49391   50092   50338   49548   50055   49486   50481   seekrandomwhilewriting
29553   27373   28491   28734   28586   28242   27853   28267   seekrandomwhilewriting

And this has the QPS from the not-cached test:

v64     v611    v612    v613    v614    v615    v616    v617    test
349918  349072  341224  347164  348470  340888  347850  334909  fillrandom
344040  327776  334852  332857  336480  343888  339678  332415  overwrite
2888170 2704291 2869560 2838685 2708847 2630220 2743535 2634374 readseq
167660  133981  130999  112923  120657  120273  87018   121126  readrandom
79615   58025   66542   49269   71643   71525   94862   71959   seekrandom
3784203 3284938 3411521 3404096 3414857 3409997 3448335 3366118 readseq
222893  165198  172372  175113  169132  174096  190337  166636  readrandom
80397   59224   67565   83540   73345   73666   94354   73855   seekrandom
78815   58232   65396   81865   72491   71938   92648   72689   seekrandom
70153   52933   60468   73907   64593   64626   80768   65317   seekrandom
36654   29389   32587   36881   34236   33753   38028   33618   seekrandom
154127  150561  150021  151168  148856  149113  149967  150643  readwhilewriting
57050   55440   55498   55576   55258   56255   55440   55162   seekrandomwhilewriting
56178   54348   55160   54893   54251   54699   54177   54770   seekrandomwhilewriting
51651   49415   50537   50627   49488   49281   49268   50172   seekrandomwhilewriting
29567   27303   28346   28359   28376   27969   27932   28204   seekrandomwhilewriting

Thursday, January 21, 2021

Summary of my search for regressions with MySQL & Postgres

This wraps up another round of benchmarks. My goal was to determine whether there were CPU regressions in MySQL and Postgres. I found more in MySQL than Postgres. 

I used the Insert Benchmark and sysbench. I also ran tests with TSBS but ran out of time to write up the results. Tests were run on small servers that I have at home. It would be nice to repeat these tests using larger servers in the cloud, perhaps someone else will do that.

The interesting result is that CPU regressions in MySQL reduce throughput by 10% to 20% between each major release. While my tests were limited to 5.6, 5.7 and 8.0 I saw similar problems years ago when I was able to compile older releases from source. Low-concurrency CPU regressions have been a thing with MySQL, and they still are a thing. Fortunately Postgres is the existence proof that the regressions can be avoided.

Sysbench results:

Insert benchmark results

Wednesday, January 20, 2021

Sysbench: MySQL vs Postgres for an IO-bound workload

This compares results for MySQL/InnoDB and Postgres via IO-bound sysbench using the data I shared for Postgres and MySQL. I previously shared a comparison for an in-memory workload with sysbench. My goal is to document how performance and efficiency have changed from MySQL 5.6 to 8.0 and from Postgres 11 to 13. I try to avoid benchmarketing but I am curious about how system software evolves over time. I share many numbers here (performance is a vector not a scalar) and am happy that one DBMS isn't strictly faster or more efficient than the other.


  • On average, old Postgres was slower than old MySQL while new Postgres is faster than new MySQL courtesy of CPU regressions in new MySQL (for this workload, HW, etc)
  • The InnoDB clustered PK index hurts throughput for a few queries when the smaller Postgres PK index can be cached, but the larger InnoDB PK index cannot.
  • The InnoDB clustered PK index helps a few queries where Postgres must to extra random IO to fetch columns not in the PK index


Much more detail on the benchmarks is here for Postgres and here for MySQL/InnoDB. The summary is that I used 1 table with 400M rows and small servers. Tests were run for Postgres (11.10, 12.4 and 13.1) and MySQL (5.6.49, 5.7.31, 8.0.22) but my focus here is comparing Postgres 11.10 with MySQL 5.6.49 and then Postgres 13.1 with MySQL 8.0.22.


The tests are in 5 groups based on the sequence in which they are run: load, read-only run before write-heavy, write-heavy, read-only run after write-heavy and insert/delete. 

I have scripts that generate 3 summaries -- absolute throughput, relative throughput and HW efficiency. Absolute throughput is the QPS or TPS for a test. Relative throughput is the QPS or TPS relative to the base case. The HW efficiency report has absolute and relative results for CPU and IO per operation.

I use ratios (relative throughput & relative HW efficiency) to explain performance. For this post the denominator (the base case) is MySQL and the numerator is Postgres. A throughput ratio < 1 means that Postgres is slower. For HW efficiency, CPU and IO per operation, a ratio > 1 means that Postgres uses more CPU or IO per operation.

I did two comparisons: one for Postgres 11 vs MySQL 5.6, another for Postgres 13 vs MySQL 8.0 and the files are in github. Files that end in old are for Postgres 11 vs MySQL 5.6 while files that end in new are for Postgres 13 vs MySQL 8.0. There are files for absolute throughput (old and new), relative throughput (old and new) and HW efficiency (old and new).

I am still figuring out how to present this data and balance both the ease of writing reports while making it not too hard to read. In this case I put the results for throughput and HW efficiency into a spreadsheet. The spreadsheet has many rows and 10 columns. The rows are the order in which the tests were run. The columns are:
  • Test - the name of the test
  • Range - the value of the range argument for Lua scripts that determines the number of rows processed for some of the tests.
  •, qps.old - throughput ratios
  • cpu/o - new, cpu/o - old  - ratios for CPU/operation
  • r/o - new, r/o - old - ratios for storage reads/operation
  • rKB/o - new, rKB/o - old - ratios for storage read KB/operation
The ratios have (Postgres value / MySQL value), new means this is for Postgres 13.1 vs MySQL 8.0.22 and old means this is for Postgres 11.10 vs MySQL 5.6.49. For the HW efficiency values (/operation) the operation is one of queries, statements, transactions.

In the spreadsheet I highlighted entries in gold when the throughput for Postgres improved by more than ~5% relative to MySQL from qps.old to (done for 19 of 43 tests). I also highlighted entries in silver when CPU/operation for Postgres improved by more than ~5% relative to MySQL from cpu/o - old to cpu/o - new (done for 26 of 43 tests). This makes it easier to understand the impact from CPU regressions.

New Postgres has improved relative to new MySQL. The geometric mean for throughput ratios is 0.95 for qps.old and 1.04 for CPU efficiency is the reason for the improvement as the geometric mean for cpu/o is 1.12 for cpu/o - old and 0.88 for cpu/o - new. Neither MySQL nor Postgres dominate on these tests, MySQL is faster for some and Postgres is faster for others. But the advantage has shifted towards Postgres thanks to CPU efficiency. On average, old Postgres was slower than old MySQL while new Postgres is faster than new MySQL courtesy of CPU regressions in new MySQL. 

There were a few outliers where MySQL is 2X faster than Postgres or vice versa and I will explain these using the names from the Test column in the spreadsheet.
  • points-covered-pk.pre (row 10) - Postgres is more than 17X faster. The PK index is covering for the query used by this test. However, InnoDB has a clustered PK index that includes all columns for the table and is much larger than the Postgres PK index. The r/o columns from the spreadsheet show that InnoDB does ~30X more storage reads/query than Postgres and that would explain the throughput difference. This also explains why Postgres is much faster for:
    • range-covered-pk.pre (row 14) - Postgres is more than 3X faster
    • points-covered-pk (row 34) - Postgres is more than 20X faster
    • range-covered-pk (row 38) - Postgres is more than 3X faster
  • update-inlist, update-nonindex (rows 18 & 20) - these do updates that don't require secondary index maintenance and MySQL gets ~2X more updates/second. MySQL modifies leaf pages of the clustered PK index while Postgres modifies both PK index leaf pages and heap-organized table pages. That might explain why Postgres uses more CPU and random IO.
  • insert (row 44) - InnoDB is ~3X faster and I suspect the change buffer is the reason
Finally, there are tests that benefit from the InnoDB clustered PK index, although the benefit there is not as extreme as the impact above where it hurts InnoDB. One example are the read-only tests (rows 8-10 and 26-28). Those queries are also included in the read-write tests (rows 24, 25) that represent the traditional sysbench test. The benefit is that InnoDB does a PK index range scan to evaluate the query predicate and get all columns, while Postgres must fetch rows via random page reads from the heap-organized table to get columns not in the PK index.