[ https://issues.apache.org/jira/browse/SPARK-22947?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16322847#comment-16322847 ]
Li Jin edited comment on SPARK-22947 at 1/11/18 7:48 PM: --------------------------------------------------------- I have looked at SPARK-8682 and tried to figure out general way to optimize range join. However, I think it's quite hard to optimize for general case of range join. Even consider the simplest form: {code:java} dfA.join(dfB, dfA.time < dfB.time) {code} For each partition of dfA [b, e) (range partition on time), we need to join it with all rows of dfB in range [b, +inf), because the predicate is unbounded. This would result in similar complexity as a cartesian product. However, in time series analysis, people usually wants to join points in time that are "close to each other". This means, the join predicate often looks like: {code:java} dfA.join(dfB, dfB.time - a < dfA.time < dfB.time + b) {code} This is bounded range predicate on a single attribute. And I think this kind of join can be optimized, described in "Optional Design Sketch: Implementation A" Then there is the second step, which is what to do with the matching rows. In the case of "as-of" join, the second step is "pick the closest row". In other operations, the second step can be "compute average" or "return all matched rows" or "return the first/last x rows". But the second step is just a local operation after potential matched data are in the same partition. This is a very common pattern in the time series analysis that we are doing and as far as I can tell, there is no good way to describe this analysis pattern in the current Spark SQL API. But I think the execution engine of Spark SQL can execute such pattern efficiently, also Catalyst can optimize such operations quite well (for instance, it's not hard for Catalyst to optimize an as-of join on "time" followed by groupby on "time" to avoid the shuffle step in groupby) Given how useful and fundamental this pattern is in time series analysis, I think it will be a good win for Spark if such pattern is supported. was (Author: icexelloss): I have looked at SPARK-8682 and tried to figured out general way to optimize range join. However, I think it's quite hard to optimize for general case of range join. Even consider the simplest form: {code:java} dfA.join(dfB, dfA.time < dfB.time) {code} For each partition of dfA [b, e) (range partition on time), we need to join it with all rows of dfB in range [b, +inf), because the predicate is unbounded. This would result in similar complexity as a cartesian product. However, in time series analysis, people usually wants to join points in time that are "close to each other". This means, the join predicate often looks like: {code:java} dfA.join(dfB, dfB.time - a < dfA.time < dfB.time + b) {code} This is bounded range predicate on a single attribute. And I think this kind of join can be optimized, described in "Optional Design Sketch: Implementation A" Then there is the second step, which is what to do with the matching rows. In the case of "as-of" join, the second step is "pick the closest row". In other operations, the second step can be "compute average" or "return all matched rows" or "return the first/last x rows". But the second step is just a local operation after potential matched data are in the same partition. This is a very common pattern in the time series analysis that we are doing and as far as I can tell, there is no good way to describe this analysis pattern in the current Spark SQL API. But I think the execution engine of Spark SQL can execute such pattern efficiently, also Catalyst can optimize such operations quite well (for instance, it's not hard for Catalyst to optimize an as-of join on "time" followed by groupby on "time" to avoid the shuffle step in groupby) Given how useful and fundamental this pattern is in time series analysis, I think it will be a good win for Spark if such pattern is supported. > SPIP: as-of join in Spark SQL > ----------------------------- > > Key: SPARK-22947 > URL: https://issues.apache.org/jira/browse/SPARK-22947 > Project: Spark > Issue Type: Improvement > Components: SQL > Affects Versions: 2.2.1 > Reporter: Li Jin > Attachments: SPIP_ as-of join in Spark SQL (1).pdf > > > h2. Background and Motivation > Time series analysis is one of the most common analysis on financial data. In > time series analysis, as-of join is a very common operation. Supporting as-of > join in Spark SQL will allow many use cases of using Spark SQL for time > series analysis. > As-of join is “join on time” with inexact time matching criteria. Various > library has implemented asof join or similar functionality: > Kdb: https://code.kx.com/wiki/Reference/aj > Pandas: > http://pandas.pydata.org/pandas-docs/version/0.19.0/merging.html#merging-merge-asof > R: This functionality is called “Last Observation Carried Forward” > https://www.rdocumentation.org/packages/zoo/versions/1.8-0/topics/na.locf > JuliaDB: http://juliadb.org/latest/api/joins.html#IndexedTables.asofjoin > Flint: https://github.com/twosigma/flint#temporal-join-functions > This proposal advocates introducing new API in Spark SQL to support as-of > join. > h2. Target Personas > Data scientists, data engineers > h2. Goals > * New API in Spark SQL that allows as-of join > * As-of join of multiple table (>2) should be performant, because it’s very > common that users need to join multiple data sources together for further > analysis. > * Define Distribution, Partitioning and shuffle strategy for ordered time > series data > h2. Non-Goals > These are out of scope for the existing SPIP, should be considered in future > SPIP as improvement to Spark’s time series analysis ability: > * Utilize partition information from data source, i.e, begin/end of each > partition to reduce sorting/shuffling > * Define API for user to implement asof join time spec in business calendar > (i.e. lookback one business day, this is very common in financial data > analysis because of market calendars) > * Support broadcast join > h2. Proposed API Changes > h3. TimeContext > TimeContext is an object that defines the time scope of the analysis, it has > begin time (inclusive) and end time (exclusive). User should be able to > change the time scope of the analysis (i.e, from one month to five year) by > just changing the TimeContext. > To Spark engine, TimeContext is a hint that: > can be used to repartition data for join > serve as a predicate that can be pushed down to storage layer > Time context is similar to filtering time by begin/end, the main difference > is that time context can be expanded based on the operation taken (see > example in as-of join). > Time context example: > {code:java} > TimeContext timeContext = TimeContext("20160101", "20170101") > {code} > h3. asofJoin > h4. User Case A (join without key) > Join two DataFrames on time, with one day lookback: > {code:java} > TimeContext timeContext = TimeContext("20160101", "20170101") > dfA = ... > dfB = ... > JoinSpec joinSpec = JoinSpec(timeContext).on("time").tolerance("-1day") > result = dfA.asofJoin(dfB, joinSpec) > {code} > Example input/output: > {code:java} > dfA: > time, quantity > 20160101, 100 > 20160102, 50 > 20160104, -50 > 20160105, 100 > dfB: > time, price > 20151231, 100.0 > 20160104, 105.0 > 20160105, 102.0 > output: > time, quantity, price > 20160101, 100, 100.0 > 20160102, 50, null > 20160104, -50, 105.0 > 20160105, 100, 102.0 > {code} > Note row (20160101, 100) of dfA is joined with (20151231, 100.0) of dfB. This > is an important illustration of the time context - it is able to expand the > context to 20151231 on dfB because of the 1 day lookback. > h4. Use Case B (join with key) > To join on time and another key (for instance, id), we use “by” to specify > the key. > {code:java} > TimeContext timeContext = TimeContext("20160101", "20170101") > dfA = ... > dfB = ... > JoinSpec joinSpec = > JoinSpec(timeContext).on("time").by("id").tolerance("-1day") > result = dfA.asofJoin(dfB, joinSpec) > {code} > Example input/output: > {code:java} > dfA: > time, id, quantity > 20160101, 1, 100 > 20160101, 2, 50 > 20160102, 1, -50 > 20160102, 2, 50 > dfB: > time, id, price > 20151231, 1, 100.0 > 20150102, 1, 105.0 > 20150102, 2, 195.0 > Output: > time, id, quantity, price > 20160101, 1, 100, 100.0 > 20160101, 2, 50, null > 20160102, 1, -50, 105.0 > 20160102, 2, 50, 195.0 > {code} > h2. Optional Design Sketch > h3. Implementation A > (This is just initial thought of how to implement this) > (1) Using begin/end of the TimeContext, we first partition the left DataFrame > intonon-overlapping partitions. For the purpose of demonstration, assume we > partition it into one-day partitions: > {code:java} > [20160101, 20160102) [20160102, 20160103) ... [20161231, 20170101) > {code} > (2) Then we partition right DataFrame into overlapping partitions, taking > into account tolerance, e.g. one day lookback: > {code:java} > [20151231, 20160102) [20160101, 20160103) ... [20161230, 20170101) > {code} > (3) Pair left and right partitions > (4) For each pair of partitions, because all data for the join is in the > partition pair, we can now join the partition pair locally. > (5) Use partitioning in (1) as the output distribution so we can reuse it for > sequential asof joins > h2. Optional Rejected Sketch > h3. Rejected Implementation A > Another implementation is to sample the data to figure out its time range > instead of using a time context. This approach is implemented in > https://github.com/twosigma/flint > This approach suffers in performance if sampling data is expensive. For > instance, when the data to be sampled is the output of an expensive > computation, sampling the data would cause the expensive computation to be > done twice. -- This message was sent by Atlassian JIRA (v6.4.14#64029) --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org