Hyperspace初体验：Delta Lake表索引

1.简介

Hyperspace是一个由微软开发的开源的数据湖索引子系统。

1.1 特性

提供了一套定义完好的索引管理API （4.1 建索引, 5.2 增量刷新）
- 为用户提供更大的自由度，毕竟用户是最了解自己用例的人
- 不尝试去解决所有问题，有些问题没有固定答案
独立于数据和元数据，索引有自己的元数据/日志（4.2 Hyperspace日志）
能感知到底层数据的版本，同数据一起“时间旅行” （5.1 Time Travel支持）
因为Hybrid Scan，索引与原始数据不需要完全同步（5.3 Hybrid Scan）

1.2 开发中

Hyperspace项目目前比较活跃，还有很多重要功能在开发中，想要预览的话可以尝试用 master 分支代码:

Spark 3.x: https://github.com/microsoft/hyperspace/issues/407
Iceberg: https://github.com/microsoft/hyperspace/issues/318
索引实现
- Z-ordering: https://github.com/microsoft/hyperspace/issues/515
- Data skipping: https://github.com/microsoft/hyperspace/issues/441

2.安装

2.1 软件版本

Hyperspace 0.4.0支持Delta Lake表的索引
Delta Lake 0.6.1是最后一个支持Spark 2.x的版本
Spark 2.4.2和Scala 2.12.8与上述Hyperspace, Delta Lake版本兼容

因为对 Apache Spark 3.x的支持还在开发中, Spark 2.4.2看起来是目前唯一能让Hyperspace和Delta Lake共同工作的Spark版本。

2.2 CLI启动

bin/spark-shell --packages io.delta:delta-core_2.12:0.6.1,com.microsoft.hyperspace:hyperspace-core_2.12:0.4.0 --conf "spark.sql.extensions=io.delta.sql.DeltaSparkSessionExtension" --conf "spark.sql.catalog.spark_catalog=org.apache.spark.sql.delta.catalog.DeltaCatalog"Ivy Default Cache set to: /Users/daichen/.ivy2/cache
The jars for the packages stored in: /Users/daichen/.ivy2/jars
:: loading settings :: url = jar:file:/Users/daichen/Software/spark-2.4.2-bin-hadoop2.7/jars/ivy-2.4.0.jar!/org/apache/ivy/core/settings/ivysettings.xml
io.delta#delta-core_2.12 added as a dependency
com.microsoft.hyperspace#hyperspace-core_2.11 added as a dependency
:: resolving dependencies :: org.apache.spark#spark-submit-parent-040a057e-2b7f-4d3a-a658-b93e6474dc47;1.0confs: [default]found io.delta#delta-core_2.12;0.6.1 in centralfound org.antlr#antlr4;4.7 in centralfound org.antlr#antlr4-runtime;4.7 in local-m2-cachefound org.antlr#antlr-runtime;3.5.2 in local-m2-cachefound org.antlr#ST4;4.0.8 in local-m2-cachefound org.abego.treelayout#org.abego.treelayout.core;1.0.3 in local-m2-cachefound org.glassfish#javax.json;1.0.4 in local-m2-cachefound com.ibm.icu#icu4j;58.2 in local-m2-cachefound com.microsoft.hyperspace#hyperspace-core_2.11;0.3.0 in central
:: resolution report :: resolve 351ms :: artifacts dl 9ms:: modules in use:com.ibm.icu#icu4j;58.2 from local-m2-cache in [default]com.microsoft.hyperspace#hyperspace-core_2.11;0.3.0 from central in [default]io.delta#delta-core_2.12;0.6.1 from central in [default]org.abego.treelayout#org.abego.treelayout.core;1.0.3 from local-m2-cache in [default]org.antlr#ST4;4.0.8 from local-m2-cache in [default]org.antlr#antlr-runtime;3.5.2 from local-m2-cache in [default]org.antlr#antlr4;4.7 from central in [default]org.antlr#antlr4-runtime;4.7 from local-m2-cache in [default]org.glassfish#javax.json;1.0.4 from local-m2-cache in [default]---------------------------------------------------------------------|                  |            modules            ||   artifacts   ||       conf       | number| search|dwnlded|evicted|| number|dwnlded|---------------------------------------------------------------------|      default     |   9   |   0   |   0   |   0   ||   9   |   0   |---------------------------------------------------------------------
:: retrieving :: org.apache.spark#spark-submit-parent-040a057e-2b7f-4d3a-a658-b93e6474dc47confs: [default]0 artifacts copied, 9 already retrieved (0kB/8ms)
21/12/09 16:15:53 WARN NativeCodeLoader: Unable to load native-hadoop library for your platform... using builtin-java classes where applicable
Using Spark's default log4j profile: org/apache/spark/log4j-defaults.properties
Setting default log level to "WARN".
To adjust logging level use sc.setLogLevel(newLevel). For SparkR, use setLogLevel(newLevel).
Spark context Web UI available at http://192.168.4.51:4040
Spark context available as 'sc' (master = local[*], app id = local-1639095359127).
Spark session available as 'spark'.
Welcome to____              __/ __/__  ___ _____/ /___\ \/ _ \/ _ `/ __/  '_//___/ .__/\_,_/_/ /_/\_\   version 2.4.2/_/Using Scala version 2.12.8 (OpenJDK 64-Bit Server VM, Java 1.8.0_312)
Type in expressions to have them evaluated.
Type :help for more information.scala>

3.Delta Lake表

3.1 建表

本文所有例子都使用Spark自带的 employees.json 数据集。首先，我们就先用这个数据集建一张Delta Lake的分区表。

scala> val employees = spark.read.json("examples/src/main/resources/employees.json")
employees: org.apache.spark.sql.DataFrame = [name: string, salary: bigint]scala> employees.show()
+-------+------+
|   name|salary|
+-------+------+
|Michael|  3000|
|   Andy|  4500|
| Justin|  3500|
|  Berta|  4000|
+-------+------+scala> employees.write.partitionBy("salary").format("delta").save("/tmp/delta-table/employees")

现在可以看一下Delta Lake的文件夹结构:

$ tree /tmp/delta-table/employees
/tmp/delta-table/employees
├── _delta_log
│   └── 00000000000000000000.json
├── salary=3000
│   └── part-00000-e0493ade-23d4-402e-bf43-0ad727b0754e.c000.snappy.parquet
├── salary=3500
│   └── part-00000-5bae7fa3-43c3-43a7-85a2-286848b5d589.c000.snappy.parquet
├── salary=4000
│   └── part-00000-090c7267-5a47-439a-8ca6-3ae670969ac3.c000.snappy.parquet
└── salary=4500└── part-00000-375f485f-d021-4625-b598-1f86b3ad8953.c000.snappy.parquet

3.2 Delta Lake日志

Delta日志记录了数据文件中所有的增加和删除操作（因为Data Lake存储，比如S3，不支持文件的更新甚至追加），这个日志本身也是 Parquet格式的:

$ cat /tmp/delta-table/employees/_delta_log/00000000000000000000.json
{"commitInfo":{"timestamp":1639095874103,"operation":"WRITE","operationParameters":{"mode":"ErrorIfExists","partitionBy":"[\"salary\"]"},"isBlindAppend":true,"operationMetrics":{"numFiles":"4","numOutputBytes":"1679","numOutputRows":"4"}}}
{"protocol":{"minReaderVersion":1,"minWriterVersion":2}}
{"metaData":{"id":"0c5bbb47-1ee3-4637-883c-61284a73501b","format":{"provider":"parquet","options":{}},"schemaString":"{\"type\":\"struct\",\"fields\":[{\"name\":\"name\",\"type\":\"string\",\"nullable\":true,\"metadata\":{}},{\"name\":\"salary\",\"type\":\"long\",\"nullable\":true,\"metadata\":{}}]}","partitionColumns":["salary"],"configuration":{},"createdTime":1639095873565}}
{"add":{"path":"salary=3000/part-00000-e0493ade-23d4-402e-bf43-0ad727b0754e.c000.snappy.parquet","partitionValues":{"salary":"3000"},"size":434,"modificationTime":1639095873996,"dataChange":true}}
{"add":{"path":"salary=3500/part-00000-5bae7fa3-43c3-43a7-85a2-286848b5d589.c000.snappy.parquet","partitionValues":{"salary":"3500"},"size":425,"modificationTime":1639095874020,"dataChange":true}}
{"add":{"path":"salary=4000/part-00000-090c7267-5a47-439a-8ca6-3ae670969ac3.c000.snappy.parquet","partitionValues":{"salary":"4000"},"size":416,"modificationTime":1639095874044,"dataChange":true}}
{"add":{"path":"salary=4500/part-00000-375f485f-d021-4625-b598-1f86b3ad8953.c000.snappy.parquet","partitionValues":{"salary":"4500"},"size":404,"modificationTime":1639095874069,"dataChange":true}}

4.Hyperspace基础

4.1 创建索引

让我们先来创建一个索引（Hyperspace目前主要支持Covering索引，即索引本身也包括其他列数据，这样查询可以直接在索引里完成，无需访问原始数据）:

scala> val employees = spark.read.format("delta").load("/tmp/delta-table/employees")
employees: org.apache.spark.sql.DataFrame = [name: string, salary: bigint]                                                                                          scala> spark.conf.set("spark.hyperspace.index.sources.fileBasedBuilders",|       "com.microsoft.hyperspace.index.sources.delta.DeltaLakeFileBasedSourceBuilder," +|       "com.microsoft.hyperspace.index.sources.default.DefaultFileBasedSourceBuilder")scala> import com.microsoft.hyperspace._
import com.microsoft.hyperspace._scala> val hyperspace = new Hyperspace(spark)
hyperspace: com.microsoft.hyperspace.Hyperspace = com.microsoft.hyperspace.Hyperspace@515a3572scala> import com.microsoft.hyperspace.index._
import com.microsoft.hyperspace.index._scala> hyperspace.createIndex(employees, IndexConfig("deltaIndex", indexedColumns = Seq("name"), includedColumns = Seq("salary")))

4.2 Hyperspace日志

$ spark-2.4.2-bin-hadoop2.7 $ tree spark-warehouse
spark-warehouse
└── indexes└── deltaIndex├── _hyperspace_log│   ├── 0│   ├── 1│   └── latestStable└── v__=0├── _SUCCESS├── part-00071-46fed65d-a011-4d69-9267-30c40c623a78_00071.c000.snappy.parquet├── part-00164-46fed65d-a011-4d69-9267-30c40c623a78_00164.c000.snappy.parquet├── part-00165-46fed65d-a011-4d69-9267-30c40c623a78_00165.c000.snappy.parquet└── part-00169-46fed65d-a011-4d69-9267-30c40c623a78_00169.c000.snappy.parquet

4.3 索引的Explain

scala> val query = employees.filter(employees("name") === "Andy").select("salary")
query: org.apache.spark.sql.DataFrame = [salary: bigint]scala> query.explain
== Physical Plan ==
*(1) Project [salary#325L]
+- *(1) Filter (isnotnull(name#324) && (name#324 = Andy))+- *(1) FileScan parquet [name#324,salary#325L] Batched: true, Format: Parquet, Location: TahoeLogFileIndex[file:/tmp/delta-table/employees], PartitionCount: 4, PartitionFilters: [], PushedFilters: [IsNotNull(name), EqualTo(name,Andy)], ReadSchema: struct<name:string>scala> query.show
+------+
|salary|
+------+
|  4500|
+------+scala> spark.enableHyperspace
res3: org.apache.spark.sql.SparkSession = org.apache.spark.sql.SparkSession@ac52e35scala> query.explain
== Physical Plan ==
*(1) Project [salary#325L]
+- *(1) Filter (isnotnull(name#324) && (name#324 = Andy))+- *(1) FileScan Hyperspace(Type: CI, Name: deltaIndex, LogVersion: 1) [name#324,salary#325L] Batched: true, Format: Parquet, Location: InMemoryFileIndex[file:/Users/daichen/Software/spark-2.4.2-bin-hadoop2.7/spark-warehouse/indexes/..., PartitionFilters: [], PushedFilters: [IsNotNull(name), EqualTo(name,Andy)], ReadSchema: struct<name:string,salary:bigint>scala> query.show
+------+
|salary|
+------+
|  4500|
+------+

除了Spark DataFrame的Explain，Hyperspace自己的Explain提供了更多细节:

scala> hyperspace.explain(query, verbose = true)
=============================================================
Plan with indexes:
=============================================================
Project [salary#325L]
+- Filter (isnotnull(name#324) && (name#324 = Andy))<----+- FileScan Hyperspace(Type: CI, Name: deltaIndex, LogVersion: 1) [name#324,salary#325L] Batched: true, Format: Parquet, Location: InMemoryFileIndex[file:/Users/daichen/Software/spark-2.4.2-bin-hadoop2.7/spark-warehouse/indexes/..., PartitionFilters: [], PushedFilters: [IsNotNull(name), EqualTo(name,Andy)], ReadSchema: struct<name:string,salary:bigint>---->=============================================================
Plan without indexes:
=============================================================
Project [salary#325L]
+- Filter (isnotnull(name#324) && (name#324 = Andy))<----+- FileScan parquet [name#324,salary#325L] Batched: true, Format: Parquet, Location: TahoeLogFileIndex[file:/tmp/delta-table/employees], PartitionCount: 4, PartitionFilters: [], PushedFilters: [IsNotNull(name), EqualTo(name,Andy)], ReadSchema: struct<name:string>---->=============================================================
Indexes used:
=============================================================
deltaIndex:file:/Users/daichen/Software/spark-2.4.2-bin-hadoop2.7/spark-warehouse/indexes/deltaIndex/v__=0=============================================================
Physical operator stats:
=============================================================
+-----------------------------------------------------------+-------------------+------------------+----------+
|                                          Physical Operator|Hyperspace Disabled|Hyperspace Enabled|Difference|
+-----------------------------------------------------------+-------------------+------------------+----------+
|*Scan Hyperspace(Type: CI, Name: deltaIndex, LogVersion: 1)|                  0|                 1|         1|
|                                              *Scan parquet|                  1|                 0|        -1|
|                                                     Filter|                  1|                 1|         0|
|                                                    Project|                  1|                 1|         0|
|                                          WholeStageCodegen|                  1|                 1|         0|
+-----------------------------------------------------------+-------------------+------------------+----------+

5.Hyperspace进阶

5.1 Time Travel支持

追加新数据

首先我们向Delta Lake表追加一条新数据，这样就会产生一个新的版本:

scala> val columns = Seq("name", "salary")
columns: Seq[String] = List(name, salary)scala> val data = Seq(("Chen", 1000L))
data: Seq[(String, Long)] = List((Chen,1000))scala> val newEmployees = spark.createDataFrame(data).toDF(columns:_*)
newEmployees: org.apache.spark.sql.DataFrame = [name: string, salary: bigint]scala> newEmployees.write.mode("append").partitionBy("salary").format("delta").save("/tmp/delta-table/employees")scala> val employees = spark.read.format("delta").load("/tmp/delta-table/employees")
employees: org.apache.spark.sql.DataFrame = [name: string, salary: bigint]scala> employees.show()
+-------+------+
|   name|salary|
+-------+------+
|Michael|  3000|
| Justin|  3500|
|  Berta|  4000|
|   Andy|  4500|
|   Chen|  1000|
+-------+------+

可以在Delta日志里看到的确有一个新的日志和数据文件产生:

$ tree /tmp/delta-table/employees
/tmp/delta-table/employees
├── _delta_log
│   ├── 00000000000000000000.json
│   └── 00000000000000000001.json
├── salary=1000
│   └── part-00000-7c9cd83b-6602-4b9a-816c-12aee7f8eccf.c000.snappy.parquet
├── salary=3000
│   └── part-00000-e0493ade-23d4-402e-bf43-0ad727b0754e.c000.snappy.parquet
├── salary=3500
│   └── part-00000-5bae7fa3-43c3-43a7-85a2-286848b5d589.c000.snappy.parquet
├── salary=4000
│   └── part-00000-090c7267-5a47-439a-8ca6-3ae670969ac3.c000.snappy.parquet
└── salary=4500└── part-00000-375f485f-d021-4625-b598-1f86b3ad8953.c000.snappy.parquet$ cat /tmp/delta-table/employees/_delta_log/00000000000000000001.json
{"commitInfo":{"timestamp":1639161717370,"operation":"WRITE","operationParameters":{"mode":"Append","partitionBy":"[\"salary\"]"},"readVersion":0,"isBlindAppend":true,"operationMetrics":{"numFiles":"1","numOutputBytes":"404","numOutputRows":"1"}}}
{"add":{"path":"salary=1000/part-00000-7c9cd83b-6602-4b9a-816c-12aee7f8eccf.c000.snappy.parquet","partitionValues":{"salary":"1000"},"size":404,"modificationTime":1639161717341,"dataChange":true}}

Time Travel

Hyperspace能感知到Delta Lake表的版本变化，并相应的采用之前版本的索引数据，这样就保证了查询结果的正确性。

scala> val oldEmployees = spark.read.format("delta").option("versionAsOf", 0).load("/tmp/delta-table/employees")
oldEmployees: org.apache.spark.sql.DataFrame = [name: string, salary: bigint]scala> oldEmployees.show()
+-------+------+
|   name|salary|
+-------+------+
|Michael|  3000|
| Justin|  3500|
|  Berta|  4000|
|   Andy|  4500|
+-------+------+scala> oldEmployees.filter(oldEmployees("name") === "Andy").select("salary").show()
+------+
|salary|
+------+
|  4500|
+------+scala> oldEmployees.filter(oldEmployees("name") === "Andy").select("salary").explain()
== Physical Plan ==
*(1) Project [salary#1180L]
+- *(1) Filter (isnotnull(name#1179) && (name#1179 = Andy))+- *(1) FileScan Hyperspace(Type: CI, Name: deltaIndex, LogVersion: 1) [name#1179,salary#1180L] Batched: true, Format: Parquet, Location: InMemoryFileIndex[file:/Users/daichen/Software/spark-2.4.2-bin-hadoop2.7/spark-warehouse/indexes/..., PartitionFilters: [], PushedFilters: [IsNotNull(name), EqualTo(name,Andy)], ReadSchema: struct<name:string,salary:bigint>

5.2 增量刷新

Hyperspace提供了全量和增量刷新索引数据的API，像下面这样增量刷新后，索引就能应用到最新的Delta Lake表了:

scala> hyperspace.refreshIndex("deltaIndex", "incremental")scala> val query = employees.filter(employees("name") === "Andy").select("salary")
query: org.apache.spark.sql.DataFrame = [salary: bigint]scala> query.explain
== Physical Plan ==
*(1) Project [salary#5066L]
+- *(1) Filter (isnotnull(name#5065) && (name#5065 = Andy))+- *(1) FileScan Hyperspace(Type: CI, Name: deltaIndex3, LogVersion: 3) [name#5065,salary#5066L] Batched: true, Format: Parquet, Location: InMemoryFileIndex[file:/Users/daichen/Software/spark-2.4.2-bin-hadoop2.7/spark-warehouse/indexes/..., PartitionFilters: [], PushedFilters: [IsNotNull(name), EqualTo(name,Andy)], ReadSchema: struct<name:string,salary:bigint>

5.3 Hybrid Scan

因为Hybrid Scan目前不支持分区表，我们首先重新创建一张无分区的Delta Lake表。

scala> val employees = spark.read.json("examples/src/main/resources/employees.json")
scala> employees.write.format("delta").save("/tmp/delta-table/employees3")
scala> val employees3 = spark.read.format("delta").load("/tmp/delta-table/employees3")scala> hyperspace.createIndex(employees, IndexConfig("deltaIndex3", indexedColumns = Seq("name"), includedColumns = Seq("salary")))
scala> newEmployees.write.mode("append").format("delta").save("/tmp/delta-table/employees3")scala> val employees3 = spark.read.format("delta").load("/tmp/delta-table/employees3")

Hybrid Scan默认是关闭的，所以我们要先开启它，同时修改应用它的最大比例。这个最大比例默认是0.3，即改动文件个数/总文件数>0.3时，Hybrid Scan不会生效。这样避免了大量索引文件过期时，强制使用Hybrid Scan可能导致的性能问题。因为我们新建的未分区表只有一个数据文件，追加一条数据产生一个新文件，此时这个比例相当于0.5，所以必须先修改才能看到Hybrid Scan的效果。

scala> spark.conf.set("spark.hyperspace.index.hybridscan.enabled", "true")
scala> spark.conf.set("spark.hyperspace.index.hybridscan.maxAppendedRatio", "1.0")

现在万事俱备，再次查询就会看到，这次不同于5.1里的现象。即便我们未刷新索引数据，它还是可以应用到最新的Delta Lake表上。

scala> val query = employees3.filter(employees3("name") === "Andy").select("salary")
query: org.apache.spark.sql.DataFrame = [salary: bigint]scala> query.explain
== Physical Plan ==
*(1) Project [salary#5571L]
+- *(1) Filter (isnotnull(name#5570) && (name#5570 = Andy))+- *(1) FileScan Hyperspace(Type: CI, Name: deltaIndex3, LogVersion: 3) [name#5570,salary#5571L] Batched: true, Format: Parquet, Location: InMemoryFileIndex[file:/Users/daichen/Software/spark-2.4.2-bin-hadoop2.7/spark-warehouse/indexes/..., PartitionFilters: [], PushedFilters: [IsNotNull(name), EqualTo(name,Andy)], ReadSchema: struct<name:string,salary:bigint>

参考引用

Spark

1.1 Apache Spark previous version download: https://archive.apache.org/dist/spark/spark-2.4.2/
1.2 Delta Lake and Spark version compatibility: https://github.com/delta-io/delta/releases
1.3 Data Frame API: https://spark.apache.org/docs/2.3.1/sql-programming-guide.html#running-sql-queries-programmatically
Delta Lake

2.1 Paper: https://databricks.com/wp-content/uploads/2020/08/p975-armbrust.pdf
2.2 Quick start: https://docs.delta.io/latest/quick-start.html#language-scala
2.3 CRUD doc: https://docs.delta.io/latest/delta-batch.html, https://docs.delta.io/latest/delta-utility.html#-delta-detail&language-scala
Hyperspace

3.1 Paper: http://vldb.org/pvldb/vol14/p3043-potharaju.pdf
3.2 Quick start: https://microsoft.github.io/hyperspace/docs/ug-quick-start-guide/
3.3 Delta Lake integration: https://microsoft.github.io/hyperspace/docs/ug-supported-data-formats/
3.4 User guide: https://docs.microsoft.com/en-us/azure/synapse-analytics/spark/apache-spark-performance-hyperspace?pivots=programming-language-scala
3.5 Hyperspace index settings: https://github.com/microsoft/hyperspace/blob/v0.4.0/src/main/scala/com/microsoft/hyperspace/index/IndexConstants.scala