how to model data / indexes to find timeslices fast - oracle

Question

We have a lot of tables in our database with data that is only relevant/valid during a certain period of time. For example contracts, they have a start_date and an end_date. And it's not necessarily full months.
Now this is a typical type of query against this table:
SELECT
*
FROM
contracts c
WHERE
c.start_date <= :1
AND c.end_date >= :2
AND c.region_id = :3
Since we have 20 years of data in our table (~7000 days), the date is very good filter criteria, especially when :1 and :2 is the same day. The region_id is not such a good filter criteria because there aren't that many (~50). In this example we have (among others) 2 Indexes on our table:
contracts_valid_index (start_date, end_date)
contracts_region (region_id)
Unfortunately, above query will often us the contracts_region index because the optimizer thinks it's cheaper. The reason behind this is simple: When I pick a day in the middle of our data, then the database will think that an index over start_date will not really be good because it will only filter out half the data. And by looking at end_date the same applies. So the optimizer thinks that he can only filter out 1/4 of my data. Because he does not know that start_date and end_date are usually pretty close together and this index would be very selective.
An execution plan using the contracts_valid_index has higher costs than an execution plan using contracts_region. But in reality the contracts_valid_index is a lot better.
I currently don't think that I can speed up my queries by making better indexes (other than deleting all but contracts_valid_index). But maybe my data model is not very good for the query optimizer. So I assume that others are also having similar needs and would love to know how they modeled their data or optimized their data tables / indexes.
Any suggestions?

Answer 1

Since you indicate you are using Oracle 12c it may help to define your Start_Date and End_Date columns as temporal valid time columns provided they match the appropriate temporal validity semantics (start_date and end_date need to be timestamps, end_date must be > start_date or possibly null and valid time periods include the start date but exclude the end date, that is it's a partially closed/open range unlike the usual between operator which denotes a fully closed range). For example:
ALTER TABLE contracts ADD (PERIOD FOR valid_time (start_date, end_date));
You can then query the contracts table for a given validity period thusly:
SELECT
c.*
FROM
contracts VERSIONS PERIOD FOR valid_time BETWEEN :1 AND :2 c
WHERE
c.region_id = :3
This is semantically similar to:
SELECT
c.*
FROM
contracts c
WHERE
:1 < end_date
AND start_date <= :2
AND c.region_id = :3
Alternatively to query for records that are valid for a specific point in time rather than a range of time:
SELECT
c.*
FROM
contracts AS OF PERIOD FOR valid_time :1 c
WHERE
c.region_id = :2
which is semantically similar to:
SELECT
c.*
FROM
contracts c
WHERE
:1 BETWEEN start_date AND end_date
and :1 <> end_date
and c.region_id = :2
I'm not sure if null values for start_date and end_date indicate the beginning and end of time respectively or not since I don't currently have an R12 instance to test in.

Answer 2

I have previously come across the same problem of index usage in relation to large sets of IP addresses on MySQL databases (bear with me; it really is the same problem).
The solution I found (by much googling, I'm not taking the credit for inventing it) was to use a geospatial index. This is specifically designed to find data within ranges. Most implementations (including that in mysql) are hard-wired to a 2 dimensional space while ip addresses and time are 1 dimensional, but its trivial to map a 1 dimensional coordinate into a 2 dimensional space (see link for a step by step explanation).
Sorry, I don't know anything about Oracle's geospatial capabilities so I can't offer any example code but, it does support geospatial indexing so can resolve your queries efficiently.

Answer 3

You could try the following query to see if it works better:
WITH t1 AS (
SELECT *
FROM contracts c
WHERE c.start_date <= :1
AND c.end_date >= :2
)
SELECT *
FROM t1
WHERE c.region_id = :3
Though it will likely prevent any possibility of using the contracts_region index.
Alternatively you could try hinting the query to use the desired index:
SELECT /*+ INDEX(c contracts_valid_index) */
*
FROM
contracts c
WHERE
c.start_date <= :1
AND c.end_date >= :2
AND c.region_id = :3
Or hinting it to not use the undesired index:
SELECT /*+ NO_INDEX(c contracts_region ) */
*
FROM
contracts c
WHERE
c.start_date <= :1
AND c.end_date >= :2
AND c.region_id = :3
When testing this out for myself without using hints I found that when selecting for dates near the start or end of the available date range the optimizer was using the INDEX_RS_ASC hint. Adding that to the query as shown below caused my testing to use the desired index even when the date range was closer to the center of the date range:
SELECT /*+ INDEX_RS_ASC(c contracts_valid_index) */
*
FROM
contracts c
WHERE
c.start_date <= :1
AND c.end_date >= :2
AND c.region_id = :3
My sample data consisted of 10,000,000 rows evenly distributed accross 50 regions and 1000 years each with a 30 day valid range.