A research cloud OLAP engine using hybrid caching and computation pushdown.

[1] Yifei Yang, Matt Youill, Matthew Woicik, Yizhou Liu, Xiangyao Yu, Marco Serafini, Ashraf Aboulnaga, Michael Stonebraker, FlexPushdownDB: Hybrid Pushdown and Caching in a Cloud DBMS, VLDB 2021.

git clone https://github.com/cloud-olap/FlexPushdownDB-Dev.git
cd FlexPushdownDB-Dev

Compiler needed:

• Linux: LLVM-12 or later versions.
• Mac OS: LLVM-13.

Set up AWS access credentials and config. Run aws configure or manually set ~/.aws/credentials and ~/.aws/config.

1. Install required dependency. For Ubuntu, ./tools/project/bin/ubuntu-prerequisites.sh. For other Linux OS or Mac OS, required dependencies listed in tools/project/bin/ubuntu-prerequisites.sh have to be manually installed.
2. Build the system ./resources/script/build.sh -s (Or using an IDE like Clion to build).

1. Create a cluster of EC2 nodes (at least 2), within which one is the coordinator, the others are executors. The coordinator is not necessarily as powerful as the executor.
2. Log in the coordinator. Put public IP of all nodes (including the coordinator) into resources/config/cluster_ips. Each line is one IP.
3. Set up the system ./resources/script/setup.sh. If all required dependencies are already installed, the system can also be set up by ./resources/script/build.sh -d, ./resources/script/deploy.sh.

• install_dependency whether to install required dependencies when running resources/script/setup.sh.
• clean whether to clean before the build starts.
• build_parallel parallel degree used for CMake build.
• build_dir_name name of the directory for the built files.
• deploy_dir_name name of the directory used for deploying to and running on a cluster.
• temp_dir_name name of the directory to store temp files after the system starts.
• pem_path the pem file (AWS private key) used to log in created EC2 nodes.

1. Make AWS access credentials and config are set.
2. Organize data files as follows on one S3 bucket:
   1. Under resources/metadata/, create a directory of which the absolute path denotes the schema name, e.g. data files under s3://tpch-sf0.01/csv/ belong to schema tpch-sf0.01/csv/.
   2. For each table T with a single partition, name the file as T.tbl for CSV format or T.parquet for Parquet format; for each table T with multiple partitions, create a directory T_sharded/ and put partition files named as T.tbl.0/T.parquet.0, T.tbl.1/T.parquet.1... into it.
3. Register data by creating metadata files under resources/metadata/. Create a directory named as the schema name, and organize metadata files as follows:
   1. schema.json For each table, specify name, format(delimiter if CSV), numPartitions, fields.
   2. stats.json For each table, specify name, rowCount, apxColumnLength(no need to be accurate, just an estimation, used by W-LFU).
   3. zoneMap.json For each table, specify name, collected min_max values of each partition on some columns, leave it empty if not collected.

1. Start the Calcite server java -jar fpdb-calcite/java/target/flexpushdowndb.thrift.calcite-1.0-SNAPSHOT.jar &.
2. cd <build directory>/fpdb-main.
3. Run tests ./fpdb-main-test -ts=<test-suite> -tc=<test-case>, available single-node test suites are ssb-sf1-single_node-no-parallel, tpch-sf0.01-single_node-no-parallel, tpch-sf0.01-single_node-parallel.
4. When finished, stop the Calcite server.

1. Make sure the security group of your nodes allows inbound and outbound traffic for TCP protocol at CAF_SERVER_PORT (4242 by default, can be changed by instructions below).
2. cd ~/FPDB-build/.
3. Start the system ./resources/script/start.sh.
4. cd fpdb-main.
5. Run tests ./fpdb-main-test -ts=<test-suite> -tc=<test-case>, available test suites are ssb-sf1-single_node-no-parallel, tpch-sf0.01-single_node-no-parallel, tpch-sf0.01-single_node-parallel, tpch-sf0.01-distributed.
6. When finished, stop the system ~/FPDB-build/resources/script/stop.sh.

The configurations are used by fpdb-main-client by default. When writing test cases, you may input your parameters in your code or read from configuration files. For example, in fpdb-main-test, parameters of execution config are specified in the code, while parameters of AWS config are read from aws.conf. In this way your test cases can be more flexible and easy to be adjusted.

• S3_BUCKET s3 bucket name where the data is in.
• SCHEMA_NAME the name of the schema to query.
• CACHE_SIZE size of local cache.
• MODE execution mode, can be one of PULLUP, PUSHDOWN_ONLY, CACHING_ONLY, HYBRID. Currently SSB tests support all modes, TPC-H tests only support PULLUP and CACHING_ONLY.
• CACHING_POLICY cache replacement policy, can be one of LRU, LFU, LFU-S (size normalized LFU), W-LFU (Weighted LFU).
• PARALLEL_DEGREE execution parallel degree of the same kind of operator (e.g. join), independent of data parallel degree which corresponds to the number of partitions of a table.
• SHOW_OP_TIMES whether to show execution times of each operator and each kind of operator.
• SHOW_SCAN_METRICS whether to show scan metrics like data load speed, selectivity, data conversion speed...
• CAF_SERVER_PORT port used by CAF for cross-node communications.
• IS_DISTRIBUTED whether the system runs across a cluster or just locally.

• S3_CLIENT_TYPE type of S3 client, can be one of S3, AIRMETTLE, MINIO. For AIRMETTLE and MINIO, access key and endpoint need to be set in fpdb-aws/src/AWSClient.cpp.
• AWS_REGION the region of the bucket that data lies in.
• NETWORK_LIMIT used to throttle data loading from S3, set to 0 if not throttling.

• SERVER_PORT port for the Calcite server.
• JAR_NAME built jar name of the Calcite server.

• SERVER_PORT port for the Calcite server, should be kept same as cpp side.
• RESOURCE_PATH absolute path of resources/ used for metadata fetching.
• exec.conf.ec2 is a fixed config for EC2 deployment and does not need to be changed.

• Replace CAF built-in communication by Flight
• Optimize Group Operator
  1. optimize internally
  2. add back the change of moving shuffle before to after parallel group

• Update the stale Parquet reader for scan operators.
• Update the stale FileScan operator to support querying local files.
• Update the stale BloomJoin and incorporate it into the physical query plan.