Optimize startup time: condition processing

[dsyer]

Benchmark data from here https://github.com/dsyer/spring-boot-startup-bench suggest that condition processing is noticeable at startup time (not dominating it, but it's non-negligible, amounting to a few hundred ms depending on hardware). The most startling evidence is the "minimal" sample app from that project where @EnableAutoConfiguration is disabled and startup time improves dramatically.

One way this could maybe be fixed is to break up spring-boot-autoconfigure into separate modules, so that only the conditions you need to be evaluated will be, for any given app, based on the classpath.

[philwebb]

Another option that we've discussed before is to somehow pre-determine @ConditionalOnClass results at compile time.

[philwebb]

See also #7939

[philwebb]

Some experimental code is here.

Running petclinic locally in the IDE with 1.5.0.RC1

2017-01-20 23:01:54.080  INFO 70090 --- [  restartedMain] o.s.s.petclinic.PetClinicApplication     : Started PetClinicApplication in 8.016 seconds (JVM running for 8.433)
2017-01-20 23:02:11.588  INFO 70118 --- [  restartedMain] o.s.s.petclinic.PetClinicApplication     : Started PetClinicApplication in 7.952 seconds (JVM running for 8.412)
2017-01-20 23:02:29.601  INFO 70143 --- [  restartedMain] o.s.s.petclinic.PetClinicApplication     : Started PetClinicApplication in 8.178 seconds (JVM running for 8.601)

And with the experimental branch (including the other optimizations we've already pushed)

2017-01-21 07:44:36.103  INFO 84153 --- [  restartedMain] o.s.s.petclinic.PetClinicApplication     : Started PetClinicApplication in 6.695 seconds (JVM running for 7.092)
2017-01-21 07:45:04.225  INFO 84192 --- [  restartedMain] o.s.s.petclinic.PetClinicApplication     : Started PetClinicApplication in 6.777 seconds (JVM running for 7.187)
2017-01-21 07:45:31.196  INFO 84232 --- [  restartedMain] o.s.s.petclinic.PetClinicApplication     : Started PetClinicApplication in 6.665 seconds (JVM running for 7.073)

[philwebb]

And with a hacked version of the benchmarks to support 1.5.0

Benchmark                                      Mode  Cnt   Score   Error  Units
ConfigServerBenchmark.devtoolsRestart          avgt   10   1.339 ± 0.114   s/op
ConfigServerBenchmark.explodedJarMain          avgt   10   3.768 ± 0.068   s/op
ConfigServerBenchmark.fatJar138                avgt   10   4.677 ± 0.062   s/op
ConfigServerBenchmark.fatJar142                avgt   10   4.506 ± 0.064   s/op
ConfigServerBenchmark.fatJar150                avgt   10   4.154 ± 0.069   s/op
ConfigServerBenchmark.fatJar150RC1             avgt   10   4.565 ± 0.041   s/op
JsaBenchmark.sharedClasses                     avgt   10   4.724 ± 0.066   s/op
JsaBenchmark.thinMain                          avgt   10   5.540 ± 0.065   s/op
MinimalBenchmark.explodedJarMain               avgt   10   2.005 ± 0.035   s/op
MinimalBenchmark.fatJar                        avgt   10   2.464 ± 0.035   s/op
PetclinicBenchmark.devtoolsRestart             avgt   10   1.987 ± 0.129   s/op
PetclinicBenchmark.exploded150JarMain          avgt   10   5.750 ± 0.112   s/op
PetclinicBenchmark.exploded150rc1JarMain       avgt   10   6.047 ± 0.097   s/op
PetclinicBenchmark.explodedJarMain             avgt   10   5.932 ± 0.067   s/op
PetclinicBenchmark.fatJar                      avgt   10   8.510 ± 0.095   s/op
PetclinicBenchmark.fatJar150                   avgt   10   8.456 ± 0.329   s/op
PetclinicBenchmark.fatJar15rc1                 avgt   10   8.763 ± 0.299   s/op
PetclinicBenchmark.noverify                    avgt   10   7.487 ± 0.079   s/op
ShadedBenchmark.explodedShadedMain             avgt   10   7.147 ± 1.496   s/op
ShadedBenchmark.shaded                         avgt   10  10.107 ± 0.962   s/op
SpringBoot138Benchmark.explodedJarLauncher     avgt   10   3.546 ± 0.178   s/op
SpringBoot138Benchmark.explodedJarMain         avgt   10   2.460 ± 0.097   s/op
SpringBoot138Benchmark.fatJar                  avgt   10   3.292 ± 0.110   s/op
SpringBoot142Benchmark.explodedJarLauncher     avgt   10   3.641 ± 0.100   s/op
SpringBoot142Benchmark.explodedJarMain         avgt   10   2.661 ± 0.043   s/op
SpringBoot142Benchmark.fatJar                  avgt   10   3.493 ± 0.142   s/op
SpringBoot150Benchmark.explodedJarLauncher     avgt   10   3.362 ± 0.092   s/op
SpringBoot150Benchmark.explodedJarMain         avgt   10   2.508 ± 0.043   s/op
SpringBoot150Benchmark.fatJar                  avgt   10   3.173 ± 0.114   s/op
SpringBoot150RC1Benchmark.explodedJarLauncher  avgt   10   3.968 ± 0.316   s/op
SpringBoot150RC1Benchmark.explodedJarMain      avgt   10   2.796 ± 0.149   s/op
SpringBoot150RC1Benchmark.fatJar               avgt   10   3.546 ± 0.054   s/op
SpringBootThinBenchmark.basic138Thin           avgt   10   2.947 ± 0.889   s/op
SpringBootThinBenchmark.basic142Thin           avgt   10   2.950 ± 0.056   s/op
SpringBootThinBenchmark.petclinicThin          avgt   10   6.399 ± 0.514   s/op

[philwebb]

RC1 comparisons:

Name	Experimental	RC1
ConfigServerBenchmark.fatJar	4.154	4.565
PetclinicBenchmark.fatJar	8.456	8.763
PetclinicBenchmark.exploded	5.750	6.047

[philwebb]

Also worth nothing that the filtering code runs in the following time:

Filtered 68 auto configuration class in 51 ms

After that's run it's pretty much as if those auto-configuration classes were never there.

[philwebb]

There's one flaw with the current implementation. We need to store @AutoConfigureBefore/@AutoConfigureAfter and @Order annotation data otherwise the sorting fails.