Sql: Optimizing BETWEEN clause

Question

I wrote a statement that takes almost an hour to run so I am asking help so I can get to do this faster. So here we go:

I am making an inner join of two tables :

<

Answer 1

There may be a very efficient way of writing this query if the intervals are deterministic because the query could be converted to an equi-join that would be amenable to more efficient hash joining.

For example if the intervals are all hourly:

ENTRY_TIME          EXIT_TIME
2000-01-15 09:00:00 2000-01-15 09:59:59
2000-01-15 10:00:00 2000-01-15 10:59:59
2000-01-15 11:00:00 2000-01-15 11:59:59
2000-01-15 12:00:00 2000-01-15 12:59:59
....

Then the join can be written as:

intervals.entry_time=trunc(measures.time,'HH')

This would reduce the cost of everything up to and including the join pretty much to a full scan of each of the tables.

However, since you have the ORDER BY operation in there, I think that a sort-merge might still beat it as the query is written right now because the optimiser will sort a smaller data set for the sort-merge than it would for the hash join (because in the latter case it would have to sort more columns of data). you could get round this by structuring the query as:

select
  measures.measure     as measure,
  measures.time        as time,
  intervals.entry_time as entry_time,
  intervals.exit_time  as exit_time
from
  intervals inner join  
  (select time, measure from measures order by time) measures
  on  intervals.entry_time=trunc(measures.time,'HH')  
/

This gives a lower cost estimate than a sort-merge on my 10.2.0.4 test instance but I'd regard it as a little risky.

So, I'd look for a sort-merge or rewrite it to allow the use of a hash join if possible.

Answer 2

You're pretty much going to get most of the rows from both tables in this case, plus you've got a sort.

The question is, does the calling process really need all the rows, or just the first few? This would change how I'd go about optimising the query.

I'll assume your calling process wants ALL the rows. Since the join predicate is not on an equality, I'd say a MERGE JOIN may be the best approach to aim for. A merge join requires its data sources to be sorted, so if we can avoid a sort the query should run as fast as it possibly can (barring more interesting approaches such as specialised indexes or materialized views).

To avoid the SORT operations on intervals and measures, you could add indexes on (measures.time,measures.measure) and (intervals.entry_time, intervals.exit_time). The database can use the index to avoid a sort, and it'll be faster because it doesn't have to visit any table blocks.

Alternatively, if you only have an index on measures.time, the query may still run ok without adding another big index - it'll run slower though because it'll probably have to read many table blocks to get the measures.measure for the SELECT clause.

Answer 3

This is quite a common problem.

Plain B-Tree indexes are not good for the queries like this:

SELECT  measures.measure as measure,
        measures.time as time,
        intervals.entry_time as entry_time,
        intervals.exit_time as exit_time
FROM    intervals
JOIN    measures
ON      measures.time BETWEEN intervals.entry_time AND intervals.exit_time
ORDER BY
        time ASC

An index is good for searching the values within the given bounds, like this:

, but not for searching the bounds containing the given value, like this:

This article in my blog explains the problem in more detail:

• Adjacency list vs. nested sets: MySQL

(the nested sets model deals with the similar type of predicate).

You can make the index on time, this way the intervals will be leading in the join, the ranged time will be used inside the nested loops. This will require sorting on time.

You can create a spatial index on intervals (available in MySQL using MyISAM storage) that would include start and end in one geometry column. This way, measures can lead in the join and no sorting will be needed.

The spatial indexes, however, are more slow, so this will only be efficient if you have few measures but many intervals.

Since you have few intervals but many measures, just make sure you have an index on measures.time:

CREATE INDEX ix_measures_time ON measures (time)

Update:

Here's a sample script to test:

BEGIN
        DBMS_RANDOM.seed(20091223);
END;
/

CREATE TABLE intervals (
        entry_time NOT NULL,
        exit_time NOT NULL
)
AS
SELECT  TO_DATE('23.12.2009', 'dd.mm.yyyy') - level,
        TO_DATE('23.12.2009', 'dd.mm.yyyy') - level + DBMS_RANDOM.value
FROM    dual
CONNECT BY
        level <= 1500
/

CREATE UNIQUE INDEX ux_intervals_entry ON intervals (entry_time)
/

CREATE TABLE measures (
        time NOT NULL,
        measure NOT NULL
)
AS
SELECT  TO_DATE('23.12.2009', 'dd.mm.yyyy') - level / 720,
        CAST(DBMS_RANDOM.value * 10000 AS NUMBER(18, 2))
FROM    dual
CONNECT BY
        level <= 1080000
/

ALTER TABLE measures ADD CONSTRAINT pk_measures_time PRIMARY KEY (time)
/

CREATE INDEX ix_measures_time_measure ON measures (time, measure)
/

This query:

SELECT  SUM(measure), AVG(time - TO_DATE('23.12.2009', 'dd.mm.yyyy'))
FROM    (
        SELECT  *
        FROM    (
                SELECT  /*+ ORDERED USE_NL(intervals measures) */
                        *
                FROM    intervals
                JOIN    measures
                ON      measures.time BETWEEN intervals.entry_time AND intervals.exit_time
                ORDER BY
                        time
                )
        WHERE   rownum <= 500000
        )

uses NESTED LOOPS and returns in 1.7 seconds.

This query:

SELECT  SUM(measure), AVG(time - TO_DATE('23.12.2009', 'dd.mm.yyyy'))
FROM    (
        SELECT  *
        FROM    (
                SELECT  /*+ ORDERED USE_MERGE(intervals measures) */
                        *
                FROM    intervals
                JOIN    measures
                ON      measures.time BETWEEN intervals.entry_time AND intervals.exit_time
                ORDER BY
                        time
                )
        WHERE   rownum <= 500000
        )

uses MERGE JOIN and I had to stop it after 5 minutes.

Update 2:

You will most probably need to force the engine to use the correct table order in the join using a hint like this:

SELECT  /*+ LEADING (intervals) USE_NL(intervals, measures) */
        measures.measure as measure,
        measures.time as time,
        intervals.entry_time as entry_time,
        intervals.exit_time as exit_time
FROM    intervals
JOIN    measures
ON      measures.time BETWEEN intervals.entry_time AND intervals.exit_time
ORDER BY
        time ASC

The Oracle's optimizer is not smart enough to see that the intervals do not intersect. That's why it will most probably use measures as a leading table (which would be a wise decision should the intervals intersect).

Update 3:

WITH    splits AS
        (
        SELECT  /*+ MATERIALIZE */
                entry_range, exit_range,
                exit_range - entry_range + 1 AS range_span,
                entry_time, exit_time
        FROM    (
                SELECT  TRUNC((entry_time - TO_DATE(1, 'J')) * 2) AS entry_range,
                        TRUNC((exit_time - TO_DATE(1, 'J')) * 2) AS exit_range,
                        entry_time,
                        exit_time
                FROM    intervals
                )
        ),
        upper AS
        (
        SELECT  /*+ MATERIALIZE */
                MAX(range_span) AS max_range
        FROM    splits
        ),
        ranges AS
        (
        SELECT  /*+ MATERIALIZE */
                level AS chunk
        FROM    upper
        CONNECT BY
                level <= max_range
        ),
        tiles AS
        (
        SELECT  /*+ MATERIALIZE USE_MERGE (r s) */
                entry_range + chunk - 1 AS tile,
                entry_time,
                exit_time
        FROM    ranges r
        JOIN    splits s
        ON      chunk <= range_span
        )
SELECT  /*+ LEADING(t) USE_HASH(m t) */
        SUM(LENGTH(stuffing))
FROM    tiles t
JOIN    measures m
ON      TRUNC((m.time - TO_DATE(1, 'J')) * 2) = tile
        AND m.time BETWEEN t.entry_time AND t.exit_time

This query splits the time axis into the ranges and uses a HASH JOIN to join the measures and timestamps on the range values, with fine filtering later.

See this article in my blog for more detailed explanations on how it works:

• Oracle: joining timestamps and time intervals

Answer 4

You can't really optimize your statement - it's pretty simple as it is.

What you could do is investigate if some indices would help you.

You're selecting on intervals.entry_time, intervals.exit_time, measures.time - are those columns indexed?

Answer 5

To summarise: your query is running against the full set of MEASURES. It matches the time of each MEASURES record to an INTERVALS record. If the window of times spanned by INTERVALS is roughly similar to the window spanned by MEASURES then your query is also running against the full set of INTERVALS, otherwise it is running against a subset.

Why that matter is because it reduces your scope for tuning, as a full table scan is the likely to be the fastest way of getting all the rows. So, unless your real MEASURES or INTERVALS tables have a lot more columns than you give us, it is unlikely that any indexes will give much advantage.

The possible strategies are:

• no indexes at all
• index on MEASURES (TIME,MEASURE)
• index on MEASURES (TIME)
• no index on MEASURES
• index on INTERVALS (ENTRY_TIME, EXIT_TIME)
• index on INTERVALS (ENTRY_TIME)
• no index on INTERVALS
• parallel query

I'm not going to present test cases for all the permutations, because the results are pretty much as we would expect.

Here is the test data. As you can see I'm using slightly larger data sets. The INTERVALS window is bigger than the MEASURES windows but not by much. The intervals are 10000 seconds wide, and the measures are taken every 15 seconds.

SQL> select min(entry_time), max(exit_time), count(*) from intervals;

MIN(ENTRY MAX(EXIT_   COUNT(*)
--------- --------- ----------
01-JAN-09 20-AUG-09       2001

SQL> select min(ts), max(ts), count(*) from measures;

MIN(TS)   MAX(TS)     COUNT(*)
--------- --------- ----------
02-JAN-09 17-JUN-09    1200001

SQL>

NB In my test data I have presumed that INTERVAL records do not overlap. This has an important corrolary: a MEASURES record joins to only one INTERVAL.

Benchmark

Here is the benchmark with no indexes.

SQL> exec dbms_stats.gather_table_stats(user, 'MEASURES', cascade=>true)

PL/SQL procedure successfully completed.

SQL> exec dbms_stats.gather_table_stats(user, 'INTERVALS', cascade=>true)

PL/SQL procedure successfully completed.

SQL> set timing on
SQL> 
SQL> select m.measure
  2         , m.ts as "TIME"
  3         , i.entry_time
  4         , i.exit_time
  5  from
  6      intervals i
  7  inner join
  8      measures m
  9      on ( m.ts between  i.entry_time and i.exit_time )
 10  order by m.ts asc
 11  /

1200001 rows selected.

Elapsed: 00:05:37.03

SQL>

MEASURES tests

Now let's build a unique index on INTERVALS (ENTRY_TIME, EXIT_TIME) and try out the various indexing strategies for MEASURES. First up, an index MEASURES TIME column only.

SQL> create index meas_idx on measures (ts)
  2  /

Index created.

SQL> exec dbms_stats.gather_table_stats(user, 'MEASURES', cascade=>true)

PL/SQL procedure successfully completed.

SQL> 
SQL> set autotrace traceonly exp
SQL> 
SQL> set timing on
SQL> 
SQL> select m.measure
  2         , m.ts as "TIME"
  3         , i.entry_time
  4         , i.exit_time
  5  from
  6      intervals i
  7  inner join
  8      measures m
  9      on ( m.ts between  i.entry_time and i.exit_time )
 10  order by m.ts asc
 11  /

1200001 rows selected.

Elapsed: 00:05:20.21

SQL>

Now, let us index MEASURES.TIME and MEASURE columns

SQL> drop  index meas_idx
  2  /

Index dropped.

SQL> create index meas_idx on measures (ts, measure)
  2  /

Index created.

SQL> exec dbms_stats.gather_table_stats(user, 'MEASURES', cascade=>true)

PL/SQL procedure successfully completed.


SQL> select m.measure
  2         , m.ts as "TIME"
  3         , i.entry_time
  4         , i.exit_time
  5  from
  6      intervals i
  7  inner join
  8      measures m
  9      on ( m.ts between  i.entry_time and i.exit_time )
 10  order by m.ts asc
 11  /

1200001 rows selected.

Elapsed: 00:05:28.54

SQL>

Now with no index on MEASURES (but still an index on INTERVALS)

SQL> drop  index meas_idx
  2  /

Index dropped.

SQL> exec dbms_stats.gather_table_stats(user, 'MEASURES', cascade=>true)

PL/SQL procedure successfully completed.

SQL> select m.measure
  2         , m.ts as "TIME"
  3         , i.entry_time
  4         , i.exit_time
  5  from
  6      intervals i
  7  inner join
  8      measures m
  9      on ( m.ts between  i.entry_time and i.exit_time )
 10  order by m.ts asc
 11  /

1200001 rows selected.

Elapsed: 00:05:24.81

SQL> 

So what difference does parallel query make ?

SQL> select /*+ parallel (4) */
  2         m.measure
  3         , m.ts as "TIME"
  4         , i.entry_time
  5         , i.exit_time
  6  from
  7      intervals i
  8  inner join
  9      measures m
 10      on ( m.ts between  i.entry_time and i.exit_time )
 11  order by m.ts asc
 12  /

1200001 rows selected.

Elapsed: 00:02:33.82


SQL>

MEASURES Conclusion

Not much difference in the elapsed time for the different indexes. I was slightly surprised that building an index on MEASURES (TS, MEASURE) resulted in a full table scan and a somewhat slower execution time. On the other hand, it is no surprise that running in parallel query is much faster. So if you have Enterprise Edition and you have the CPUs to spare, using PQ will definitely reduce the elapsed time, although it won't change the resource costs much (and actually does a lot more sorting).

INTERVALS tests

So what difference might the various indexes on INTERVALS make? In the following tests we will retain an index on MEASURES (TS). First of all we will drop the primary key on both INTERVALS columns and replace it with a constraint on INTERVALS (ENTRY_TIME) only.

SQL> alter table intervals drop  constraint ivl_pk drop  index
  2  /

Table altered.

SQL> alter table intervals add  constraint ivl_pk primary key (entry_time) using  index
  2  /

Table altered.

SQL> exec dbms_stats.gather_table_stats(user, 'INTERVALS', cascade=>true)

PL/SQL procedure successfully completed.


SQL> select m.measure
  2         , m.ts as "TIME"
  3         , i.entry_time
  4         , i.exit_time
  5  from
  6      intervals i
  7  inner join
  8      measures m
  9      on ( m.ts between  i.entry_time and i.exit_time )
 10  order by m.ts asc
 11  /

1200001 rows selected.

Elapsed: 00:05:38.39

SQL> 

Lastly with no index on INTERVALS at all

SQL> alter table intervals drop  constraint ivl_pk drop  index
  2  /

Table altered.

SQL> exec dbms_stats.gather_table_stats(user, 'INTERVALS', cascade=>true)

PL/SQL procedure successfully completed.

SQL> select m.measure
  2         , m.ts as "TIME"
  3         , i.entry_time
  4         , i.exit_time
  5  from
  6      intervals i
  7  inner join
  8      measures m
  9      on ( m.ts between  i.entry_time and i.exit_time )
 10  order by m.ts asc
 11  /

1200001 rows selected.

Elapsed: 00:05:29.15

SQL> 

INTERVALS conclusion

The index on INTERVALS makes a slight difference. That is, indexing (ENTRY_TIME, EXIT_TIME) results in a faster execution. This is because it permist a fast full index scan rather than a full table scan. This would be more significant if the time window delineated by INTERVALS was considerably wider than that of MEASURES.

Overall Conclusions

Because we are doing full table queries none of the indexes substantially changed the execution time. So if you have Enterprise Edition and multiple CPUs Parallel Query will give you the best results. Otherwise the most best indexes would be INTERVALS(ENTRY_TIME, EXIT_TIME) and MEASURES(TS) . The Nested Loops solution is definitely faster than Parallel Query - see Edit 4 below.

If you were running against a subset of MEASURES (say a week's worth) then the presence of indexes would have a bigger impact, It is likely that the two I recommended in the previous paragraph would remain the most effective,

Last observation: I ran this on a bog standard dual core laptop with an SGA of just 512M. Yet all of my queries took less than six minutes. If your query really takes an hour then your database has some serious problems. Although this long running time could be an artefact of overlapping INTERVALS, which could result in a cartesian product.

**Edit **

Originally I included the output from

SQL> set autotrace traceonly stat exp

But alas SO severely truncated my post. So I have rewritten it but without execution or stats. Those who wish to validate my findings will have to run the queries themselevs.

Edit 4 (previous edit's removed for reasons of space)

At the third attempt I have been able to reproduce teh performance improvement for Quassnoi's solution.

SQL> set autotrace traceonly stat exp
SQL>
SQL> set timing on
SQL>
SQL> select
  2          /*+ LEADING (i) USE_NL(i, m) */
  3              m.measure
  4             , m.ts as "TIME"
  5             , i.entry_time
  6             , i.exit_time
  7  from
  8      intervals i
  9  inner join
 10      measures m
 11      on ( m.ts between  i.entry_time and i.exit_time )
 12  order by m.ts asc
 13  /

1200001 rows selected.

Elapsed: 00:00:18.39

Execution Plan
----------------------------------------------------------
Plan hash value: 974071908

---------------------------------------------------------------------------------------------------
| Id  | Operation                     | Name      | Rows  | Bytes |TempSpc| Cost (%CPU)| Time     |
---------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT              |           |  6003K|   257M|       |   973K  (1)| 03:14:46 |
|   1 |  SORT ORDER BY                |           |  6003K|   257M|   646M|   973K  (1)| 03:14:46 |
|   2 |   NESTED LOOPS                |           |       |       |       |            |          |
|   3 |    NESTED LOOPS               |           |  6003K|   257M|       |   905K  (1)| 03:01:06 |
|   4 |     TABLE ACCESS FULL         | INTERVALS |  2001 | 32016 |       |  2739   (1)| 00:00:33 |
|*  5 |     INDEX RANGE SCAN          | MEAS_IDX  | 60000 |       |       |   161   (1)| 00:00:02 |
|   6 |    TABLE ACCESS BY INDEX ROWID| MEASURES  |  3000 | 87000 |       |   451   (1)| 00:00:06 |
---------------------------------------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------

   5 - access("M"."TS">="I"."ENTRY_TIME" AND "M"."TS"<="I"."EXIT_TIME")


Statistics
----------------------------------------------------------
         66  recursive calls
          2  db block gets
      21743  consistent gets
      18175  physical reads
          0  redo size
   52171689  bytes sent via SQL*Net to client
     880416  bytes received via SQL*Net from client
      80002  SQL*Net roundtrips to/from client
          0  sorts (memory)
          1  sorts (disk)
    1200001  rows processed

SQL> 

So Nested Loops are definitely the way to go.

Useful lessons from the exercise

1. Running diagnostic tests is far more valuable than guessing and theorising
2. Understanding the data is crucial
3. Even with 11g we still soemtimes need to use hints to prod the optimizer in certain cases

Answer 6

try a parallel query

  alter session enable parallel query;

  select /*+ parallel */ ... same as before;

You could also create a materialized view, perhaps with the parallel hint above. It may take a long time to create the MV but once created it can be queried repeatedly.