dremio 属于一个比较复杂的系统，官方有不少模块，官方同时也包含了一个不错的单元测试可以基本了解查询的处理
从session到查询，到sql 解析，关系节点转换，逻辑计划器以及物理计划以及执行计划
参考代码
public class Limit0LogicalToPhysicalTest extends BaseTestQuery {
private static String TEST_PATH = TestTools.getWorkingPath() + "/src/test/resources";
private static File tblPath = null;
@BeforeClass
public static void createTable() throws Exception {tblPath = new File(getDfsTestTmpSchemaLocation(), "yelp");
FileUtils.deleteQuietly(tblPath);
FileUtils.copyFileToDirectory(new File(TEST_PATH + "/yelp_business.json"), tblPath);
FileUtils.moveFile(new File(tblPath + "/yelp_business.json"), new File(tblPath + "/1.json"));
FileUtils.copyFile(new File(tblPath + "/1.json"), new File(tblPath + "/2.json"));
}
@AfterClass
public static void cleanUpTable() throws Exception {FileUtils.deleteQuietly(tblPath);
}
@Ignore
public void ExchangesKeepTest() throws Exception {
final String yelpTable = TEMP_SCHEMA + ".\"yelp\"";
final String sql = "SELECT nested_0.review_id AS review_id, nested_0.user_id AS user_id, nested_0.votes AS votes," +
" nested_0.stars AS stars, join_business.business_id AS business_id0, join_business.neighborhoods AS neighborhoods, join_business.city AS city, join_business.latitude AS latitude, join_business.review_count AS review_count, join_business.full_address AS full_address, join_business.stars AS stars0, join_business.categories AS categories, join_business.state AS state, join_business.longitude AS longitude\n" +
"FROM (\n" +
" SELECT review_id, user_id, votes, stars, business_id\n" +
" FROM cp.\"yelp_review.json\" where 1 = 0\n" +
") nested_0\n" +
" FULL JOIN " + yelpTable + " AS join_business ON nested_0.business_id = join_business.business_id";
// 上下文配置信息
final SabotContext context = getSabotContext();
context.getOptionManager().setOption(
OptionValue.createLong(OptionValue.OptionType.SYSTEM, "planner.slice_target", 1)
);
context.getOptionManager().setOption(
OptionValue.createLong(OptionValue.OptionType.SYSTEM, "planner.width.max_per_node", 10)
);
context.getOptionManager().setOption(
OptionValue.createBoolean(OptionValue.OptionType.SYSTEM, "planner.enable_mux_exchange", true)
);
// 查询上下文
final QueryContext queryContext = new QueryContext(session(), context, UserBitShared.QueryId.getDefaultInstance());
// AttemptObserver
final AttemptObserver observer = new PassthroughQueryObserver(ExecTest.mockUserClientConnection(null));
// SqlConverter 进行sql 转换处理的
final SqlConverter converter = new SqlConverter(
queryContext.getPlannerSettings(),
queryContext.getOperatorTable(),
queryContext,
queryContext.getMaterializationProvider(),
queryContext.getFunctionRegistry(),
queryContext.getSession(),
observer,
queryContext.getCatalog(),
queryContext.getSubstitutionProviderFactory(),
queryContext.getConfig(),
queryContext.getScanResult(),
queryContext.getRelMetadataQuerySupplier());
// 解析sqlnode
final SqlNode node = converter.parse(sql);
// sql 处理器配置
final SqlHandlerConfig config = new SqlHandlerConfig(queryContext, converter, observer, null);
// sql 校验以及转换，依赖catalog，实际上是基于了calcite 处理
final ConvertedRelNode convertedRelNode = PrelTransformer.validateAndConvert(config, node);
final RelDataType validatedRowType = convertedRelNode.getValidatedRowType();
final RelNode queryRelNode = convertedRelNode.getConvertedNode();
final Rel drel = PrelTransformer.convertToDrel(config, queryRelNode, validatedRowType);
// 物理节点获取
final Pair<Prel, String> convertToPrel = PrelTransformer.convertToPrel(config, drel);
final Prel prel = convertToPrel.getKey();
final String prePhysicaltextPlan = convertToPrel.getValue();
assertThat(prePhysicaltextPlan).contains("HashToRandomExchange");assertThat(prePhysicaltextPlan).contains("UnorderedMuxExchange");assertThat(prePhysicaltextPlan).contains("Empty");assertThat(prePhysicaltextPlan).contains("EasyScan");// 物理操作器
final PhysicalOperator pop = PrelTransformer.convertToPop(config, prel);
// 物理计划
final PhysicalPlan plan = PrelTransformer.convertToPlan(config, pop);
final String postPhysicaltextPlan = plan.unparse(config.getContext().getLpPersistence().getMapper().writer());
assertThat(postPhysicaltextPlan).contains("EmptyValues");assertThat(postPhysicaltextPlan).contains("EasyGroupScan");assertThat(postPhysicaltextPlan).contains("unordered-mux-exchange");assertThat(postPhysicaltextPlan).contains("hash-to-random-exchange");
PhysicalPlanReader pPlanReader = new PhysicalPlanReader(
DEFAULT_SABOT_CONFIG, CLASSPATH_SCAN_RESULT, new LogicalPlanPersistence(DEFAULT_SABOT_CONFIG, CLASSPATH_SCAN_RESULT),
CoordinationProtos.NodeEndpoint.getDefaultInstance(),
DirectProvider.wrap(Mockito.mock(CatalogService.class)), context);
// 执行计划
ExecutionPlan exec = ExecutionPlanCreator
.getExecutionPlan(queryContext, pPlanReader, AbstractMaestroObserver.NOOP, plan,
QueueType.SMALL);
List<PlanFragmentFull> fragments = exec.getFragments();
int scanFrags = 0;
for (PlanFragmentFull fragment : fragments) {if (new String(fragment.getMajor().getFragmentJson().toByteArray()).contains("easy-sub-scan")) {scanFrags++;
}
}
assertEquals(2, scanFrags);
}
// 用户会话创建，mock 使用
private static UserSession session() {return UserSession.Builder.newBuilder()
.withSessionOptionManager(
new SessionOptionManagerImpl(getSabotContext().getOptionValidatorListing()),
getSabotContext().getOptionManager())
.withUserProperties(UserProtos.UserProperties.getDefaultInstance())
.withCredentials(UserBitShared.UserCredentials.newBuilder().setUserName("foo").build()).setSupportComplexTypes(true)
.build();
}
}
说明
以上是一个简单的说明，实际上dremio 执行是比较复杂的，使用了不同的优化器，同时还包含了物化处理，查询规则，底层数据存储读取，元数据关联
以上是一个单元测试的，实际执行还包含了具体命令的处理，线程调度。。。。。，但是基于上边可以大致了解执行的处理是值得参考的
参考资料
sabot/kernel/src/test/java/com/dremio/exec/planner/sql/handlers/commands/Limit0LogicalToPhysicalTest.java
