Awesome!!!!!!

Nice job Ethan Wang!

Thanks James Taylor!

FAILURE: Integrated in Jenkins build Phoenix-master #1725 (See https://builds.apache.org/job/Phoenix-master/1725/)
PHOENIX-153 Implement TABLESAMPLE clause (Ethan Wang) (jamestaylor: rev 5e33dc12bc088bd0008d89f0a5cd7d5c368efa25)

• (edit) phoenix-core/src/main/java/org/apache/phoenix/parse/ParseNodeFactory.java
• (edit) phoenix-core/src/main/java/org/apache/phoenix/optimize/QueryOptimizer.java
• (edit) phoenix-core/src/main/java/org/apache/phoenix/parse/FilterableStatement.java
• (edit) phoenix-core/src/main/java/org/apache/phoenix/iterate/BaseResultIterators.java
• (add) phoenix-core/src/main/java/org/apache/phoenix/iterate/TableSamplerPredicate.java
• (add) phoenix-core/src/it/java/org/apache/phoenix/end2end/QueryWithTableSampleIT.java
• (edit) phoenix-core/src/main/java/org/apache/phoenix/compile/JoinCompiler.java
• (edit) phoenix-core/src/main/java/org/apache/phoenix/parse/SelectStatement.java
• (edit) phoenix-core/src/main/java/org/apache/phoenix/parse/DeleteStatement.java
• (edit) phoenix-core/src/main/java/org/apache/phoenix/parse/ConcreteTableNode.java
• (edit) phoenix-core/src/main/antlr3/PhoenixSQL.g
• (edit) phoenix-core/src/main/java/org/apache/phoenix/parse/NamedTableNode.java

Pushed to master and 4.x branches. Great first commit, Ethan Wang!

+1. Nice work, Ethan Wang. Will let you know if patch doesn't apply cleanly to other 4.x branches.

Github user aertoria commented on the issue:

https://github.com/apache/phoenix/pull/262

Commit 0507d4f change list:
1, explain plan a new way (Thanks for the suggestion)
2, squash previous four commit into one
3, revise all commit message to start with PHOENIX-153+space

preview on a Single select
`CLIENT 3-CHUNK 30 ROWS 2370 BYTES PARALLEL 1-WAY 0.2-SAMPLED ROUND ROBIN FULL SCAN OVER PERSON`

on a Join select
```
CLIENT 1-CHUNK 0 ROWS 0 BYTES PARALLEL 1-WAY 0.65-SAMPLED ROUND ROBIN FULL SCAN OVER INX_ADDRESS_PERSON
SERVER FILTER BY FIRST KEY ONLY
PARALLEL INNER-JOIN TABLE 0
CLIENT 1-CHUNK 1 ROWS 32 BYTES PARALLEL 1-WAY 0.15-SAMPLED ROUND ROBIN FULL SCAN OVER US_POPULATION
DYNAMIC SERVER FILTER BY TO_CHAR("INX_ADDRESS_PERSON.0:ADDRESS") IN (US_POPULATION.STATE)
```

Make sense. Thanks James Taylor

Seems like review comments aren't appearing here in JIRA (maybe because your commit message doesn't include the JIRA number in the expected format), so I'll repeat it here:

Let's move the explain for the sampling into the first line, before we recurse down for the other steps. You can put it on the same line, after the "-WAY " like this:

CLIENT PARALLEL 1-WAY 0.48-SAMPLED ...

Otherwise, users will interpret the sampling as happening after the scan/filtering which isn't the case.

Github user JamesRTaylor commented on the issue:

https://github.com/apache/phoenix/pull/262

Ping @aertoria - would you have a few spare cycles to make that last change? Also, please squash all commits into one and amend your commit message to be prefixed with PHOENIX-153 (i.e. include the dash). Otherwise, we the pull request isn't tied to the JIRA.

The default guidepost width is 300MB. Maybe we could go down to 10MB, once we have guidepost combining.
Less than that will be a huge management burden to the system.

Still a good thing to do! On small tables you do not need to sample in the first place, and for large tables - where it matters - we'll have sufficiently many guide posts. (A 1TB table has over 3000 300MB guideposts, i.e. you'll have a resolution of 0.03%, which is plenty good!)

Valid Point.

In addition, by design, this coarse problem gets magnified when three things happen (and vice versa):
1, Table is too small
2, Guidepost width set too wide, or even no stats collected at all
3, User specifies to not use stats table for parallelization.

Based on the observation from the testing on a table with 400K rows and GUIDE_POSTS_WIDTH =10KB or 200KB, the sampled size was usually around +-5% of expected size. This performance gets better and better when the GuidePosts used are more granular (Detailed chart attached.)

Note that in this chart,
1, The test environment is a single node hbase cluster (1.2.2). Test table with random integer as PK, with about 400K rows.
2, The Guide_Post_Width has been pre-set as 10K and 200K, respectively.
3, FNV is used as the consistent hashing algorithm as TableSamplerPredicate (a.k.a., the "dice", used to hashing each guidepost and decide if it is going to be selected).

Good idea. Skipping whole guideposts is pretty coarse, though.
At the same time I cannot thing of anything else efficient.

Yes, +1 to following Calcite syntax

+1. After some study about calcite/parse.jj and calcite/SqlValidatorFeatureTest.java, my understanding is that calcite seems to be very close to Postgres TABLESAMPLE syntax (which PHOENIX-153 is also designed to be similar with).

I'd like to sum up two differences below (please correct me if I'm mistaken Julian Hyde).

1, Calcite table sampling rate input is 0 to 100 (PHOENIX-153 currently is 0 to 1).
2, Syntax difference
Calcite: select name from dept TABLESAMPLE system(58)
PHOENIX-153: select name from dept TABLESAMPLE 0.58

Purposing change for PHOENIX-153: Let's change phoenix side to be
select name from dept TABLESAMPLE(0.58)

Thoughts?

Reference:
https://github.com/apache/calcite/blob/d619304070bf2874ab760c92ec2573ee6c19f536/piglet/src/main/javacc/PigletParser.jj

https://github.com/apache/calcite/blob/0938c7b6d767e3242874d87a30d9112512d9243a/core/src/test/java/org/apache/calcite/test/SqlValidatorFeatureTest.java

+1. Do you have something you can point us to for the Calcite TABLESAMPLE syntax?

Since Calcite already supports TABLESAMPLE let's save ourselves a headache and make sure that the 4.x syntax is compatible with Calcite's syntax.

P.R link

https://github.com/apache/phoenix/pull/262

Spec of this patch. Feedback plz.

Updated Jun25. Syntax made similar to Calcite and PostgresSQL

++
++Belows are SUPPORTED
++
===BASE CASE====
select * from Person;
select * from PERSON TABLESAMPLE(45);
select * from PERSON TABLESAMPLE (45);

===WHERE CLAUSE====
select * from PERSON where ADDRESS = 'CA' OR name>'tina3';
select * from PERSON TABLESAMPLE (49) where ADDRESS = 'CA' OR name>'tina3';
select * from PERSON TABLESAMPLE (49) where ADDRESS = 'CA' OR name>'tina3' ORDER BY ADDRESS LIMIT 2;

===Wired Table===
select * from LOCAL_ADDRESS TABLESAMPLE (79);
select * from SYSTEM.STATS TABLESAMPLE (41);

===CORNER CASE===
select * from PERSON TABLESAMPLE (0);
select * from PERSON TABLESAMPLE (100.0);
select * from PERSON TABLESAMPLE (145.99); (out of boundary exception)
select * from PERSON TABLESAMPLE ko; (syntax error exception)

===AGGREGATION===
select count( * ) from PERSON TABLESAMPLE (49) LIMIT 2
select count( * ) from (select NAME from PERSON TABLESAMPLE (49) limit 20)

===SUB QUERY===
select * from (select /+NO_INDEX/ * from PERSON tablesample (10) where Name > 'tina10') where ADDRESS = 'CA'

===JOINS===
select * from PERSON1, PERSON2 tablesample (70) where PERSON1.Name = PERSON2.NAME
select /+NO_INDEX/ count( * ) from PERSON tableSample (19), US_POPULATION tableSample (28) where PERSON.Name > US_POPULATION.STATE

===QUERY being OPTMIZED===
select * from PERSON tablesample (80) (may go to IDX_ADDRESS_PERSON index table, but table sampling rate carry on)

===INSERT SELECT====
upsert into personbig(ID, ADDRESS) select id, address from personbig tablesample (1);

Implementation Proposal. (Feedback plz)
Proposing table sampling on each Table basis (at the 'FROM' part of the query). Sample size decided by the input sampling rate applies on Primary Key's frequency.

Syntax:
`select name from person SAMPLE(0.10) where name='ethan'`

Returns:
`person SAMPLE(0.10)` part returns rows about 10% volume of the PERSON table. Reducing performance cost from PERSON table scan to Person-STATS table scan.

Implementation detail:
For table PERSON, assume STATS is populated with GuidePost inserted on every other PK (50% coverage).
Step1, within the query scanning keyrange, iterate through the STATS table.
Step2, for every GuidePost encountered, consult with a random number generator to decide if this guidepost will be included or excluded from the sample. This dice has 10% chance of winning.
Step3, Once we decide to include this GuidePost, every PK on the original PERSON table that is between this-GuidePost and next-GuidePost will be included to the final sample.
Repeat this process untill all GuidePost are visited.

Example:

PERSON

STATS

During dice rolling process, GuidePost 3 is included. PK between [3,5) will be included. The final result will be rows with PK 3, 4.

This implementation,
a, similar to Microsoft SQLServer TABLESAMPLE, focus mainly on the performance benefit. It does not guarantee the even distribution of the sample on original table (representativity).
b, it works well on any GUIDE_POST_SWIDTH on any input sample rate. However, if the table is too small, the sample output may include rows more or less than the expected count (sample_rate X table_size)

--------------------------------------------------------------------------------

Summary of other popular TABLESAMPLE implementations.
Basically two categories:
1, Sampling on Query Basis.
(Such as Blink DB. https://sameeragarwal.github.io/blinkdb_eurosys13.pdf)
This implementation places sampling process based on entire query. such as:
`select name from person where name='ethan' SAMPLE WITH ERROR 10% CONFIDENCE 95%'

BlinkDB did so by assuming "the data used for similar grouping and filtering clause does not change over time across future queries". Based on heuristic experience, query engine pre-build certain stratify sample groups extracted from the actual table, cache them, and use them for evaluating an approximate result for some expensive queries. Therefore to avoid full table scan.

This approach:
a, Optimizes for the best performance-accuracy-trade-off. Once given the accuracy tolerance, it automatically decide the sampling rate for user.
b, Engine takes filtering and grouping into consideration therefore it's powerful. But on the other side it may not perform at the same level for all kinds of queries.
c, Based on heuristic info, there will be a machine gradually learning process.

2, Sampling on Table Basis.
(Such as Postgres, MS SQLServer. https://wiki.postgresql.org/wiki/TABLESAMPLE_Implementation)
This approach places Tablesample only on the "FROM" part of the query. such as:
`select name from person TABLESAMPLE(10 PERCENT) where name='ethan'`

This approach first sample the original table to a smaller 'view' based on the Primary Key frequency and a given sampling rate. Then that 'view' will participate into the rest part of the query in place of original table.

Usually a randomly selection process is used during the view creation. In MS SQLServer, a linear one-pass pointer travel through each "page", and ask a random generator to decide if this page will be part of the sample. Once accepted, every single row on this page now become part of new sample.

This MSSQL tablesample
a, gives flexibility satisfying any sampling rate.
b, gain performance by reducing the length of a table scan (but big O complexity still the same)
c, only care about the performance gain, does't care about sample distribution.

James Taylor Geoffrey Jacoby
Samarth Jain

Another potential means of doing sampling (as pointed out by Lars Hofhansl) is when a table is salted. Given that a salt is limited to a single byte, this may not work that well. However, if multiple bytes could be specified (PHOENIX-3574), that would be a good option.