7.6 KiB
HikariCP Ultimate JDBC Connection Pool We came, we saw, we kicked its ass
There is nothing faster. There is nothing more reliable. There is nothing more correct.
Are you using DBCP, C3P0 or BoneCP? Stop.
TL;DR
Let's look at some performance numbers. HikariCP was only compared to BoneCP because, really, DBCP and C3P0 are old and slow and I don't know why anyone would use them.
MixedBench
This is the so called "Mixed" benchmark, and it executes a representative array of JDBC operations in a realistic mix. We think median is the number to pay attention to, rather than average (which can get skewed). Median meaning 50% of the iterations were slower, %50 were faster. 200 threads were started, and the underlying connection pool contained 100 connections.
| Pool | Med (ms) | Avg (ms) | Max (ms) |
|---|---|---|---|
| BoneCP | 2155 | 1541 | 3265 |
| HikariCP | 230 | 139 | 526 |
A breakdown of the mix operations is:
| Operation | Invocations |
|---|---|
| DataSource.getConnection() | 1000 |
| PreparedStatement.prepareStatement() | 200,000 |
| PreparedStatement.setInt() | 30,000,000 |
| PreparedStatement.addBatch() | 10,000,000 |
| PreparedStatement.executeBatch() | 100,000 |
| PreparedStatement.executeQuery() | 100,000 |
| PreparedStatement.close() | 200,000 |
| ResultSet.next() | 10,000,000 |
| ResultSet.getInt() | 10,000,000 |
| ResultSet.close() | 100,000 |
| Connection.close() | 1000 |
The JVM JIT was "warmed up" with a single run through, then 4 runs were made from which the run with the lowest median time was chosen. The run with the lowest median was also the run with the lowest average -- so don't think we're skewing the results.
The benchmark was run using a stub (nop) implementation of an underlying DataSource, Connection, PreparedStatement, and ResultSet, so the driver was taken completely out of the equation so that the performance and overhead of the pools themselves could be measured.
The test was performed on an Intel Core i7 (3770 Ivy Bridge) 3.4GHz iMac with 32GB of RAM. The
JVM benchmark was run with: -server -XX:+UseParallelGC -Xss256k.
In Summary
200 threads ran 60,702,000 JDBC operations each, HikariCP did this in a median of 230ms per thread.
(In)correctness
Sometimes "correctness" is objective, and sometimes it is subjective. One example of
objective incorrectness in BoneCP is ResultSet handling. Every connection pool needs to
wrap the underlying Connection, Statement, CallableStatement, and
PreparedStatement, and ResultSet classes. However, BoneCP does not wrap ResultSet.
ResultSet must be wrapped, because ResultSet.getStatement() must return the wrapped
Statement that generated it, not the underlying Statement. Hibernate 4.3 for one
relies on this semantic.
If BoneCP were to wrap ResultSet, which comprises 20,100,000 of the 60,702,000 operations in MixedBench, its performance numbers would be far poorer. Also take note that HikariCP does properly wrap ResultSet and still achives the numbers above.
One example of subjective incorrectness -- being my personal opinion -- is that
BoneCP does not test a Connection immediately before dispatching it from the pool. It is
through this mechanism that it achives some of it's speed. In my opinion, this one "flaw"
(or "feature") renders BoneCP insuitable for Production use. The number one responsibility of
a connection pool is to not give out possibly bad connections. If you have ever run a
load-balancer in front of read-slaves, or have ever needed to bounce the DB while the
application was running, you certainly didn't do it with BoneCP.
Over on the BoneCP site, you can find a comparison of BoneCP vs. DBCP and C3P0. DBCP and C3P0,
as poor as they are, at least are performing aliveness tests before dispatching connections.
So, it's not really a fair comparison. HikariCP supports the JDBC4 Connection.isValid()
API, which for many drivers provides a non-query based aliveness test.
A particularly silly "benchmark" on the BoneCP site starts 500 threads each performing 100 DataSource.getConnection() / connection.close() calls with 0ms delay between. Who does that? The typical "mix" is dozens or hundreds of JDBC operations between obtaining the connection and closing it (hence the "MixBench") above. But ok, we can run this "benchmark" too (times in Microseconds):
| Pool | Med (μs) | Avg (μs) | Max (μs) |
|---|---|---|---|
| BoneCP | 19467 | 8762 | 30851 |
| HikariCP | 76 | 65 | 112 |
The times are per-thread reflecting 100 getConnection()/close() operations with no wait between.
Knobs
Where are all the knobs? You know, the ones used to tune and tweak the connection pool?
HikariCP has plenty of "knobs" as you can see in the configuration section below, but
comparatively less than some other pools. This was a design decision, not lack of development resources or foresight.
Missing Knobs
Some knobs and features were intentionally left out. Here are the reasons.
Statement Cache
Many (most?) major database JDBC drivers already have a PreparedStatement cache that can be
configured (Oracle, MySQL, PostgreSQL, Derby, etc). A cache in the pool would add unneeded
weight, no additional functionality, and possibly incorrect behavior...
JDBC drivers have a special relationship with the remote database in that they are directly connected and can share internal state that is synchronized with the backend. It is inherently unsafe to cache PreparedStatements outside of the driver. Why? Again, drivers have a special relationship with the database.
Take for example DDL, this is from a real world application I encountered (using BoneCP btw). Data was inserted into a table lets call X1. Every night, programatically X1 was renamed to X2, and a new X1 was created with identical structure. Basically, the application was "rolling" tables over daily (while running). In spite of the structure of X1 being identical after rolling, the database considered a PreparedStatement compiled against the old table to be invalid (probably there were some kind of UUID contained within). When the statement pool returned one of these statements, the application blew up. Turning off the cache in the connection pool, and enabling it in the driver fixed the issue. How? It is only speculation, but it seems likely that driver in this case checks with the DB to ensure the statement is still valid and if not recompiles it transparently.
Regardless of correctness or not (I'm sure it varies by DB vendor), it is unnecessary with modern databases.
Log Statement Text
Like Statement caching, most major database vendors support statement logging through
properties of their own driver. This includes Oracle, MySQL, Derby, MSSQL. For those few
databases that do not support it, log4jdbc is a good option. Actually log4jdbc provides some nice additional stuff like various timings, and
PreparedStatement bound parameter logging.
HikariCP is focused (primarily) on performance in a Production environment, so it is doubtful that we will ever implement this feature given inherent driver support and alternative solutions.