Re: Compressing large column by moving it to a unique table
On original_hashed, I think I'd try moving start_timestamp into its own
index.
On Tue, Feb 13, 2024 at 12:02 PM Adrian Garcia Badaracco <
adr...@adriangb.com> wrote:
> I'm using PostgreSQL 15.5.
>
> Here's a self-contained example. I included links to public pgMustard
> query plans.
>
> Gist link:
> https://gist.github.com/adriangb/05a01cca99a438de2a286945903c44f4
>
> Also copied below for archiving:
>
> ```sql
> CREATE OR REPLACE FUNCTION random_bytes(length integer)
> RETURNS bytea AS $$
> DECLARE
> bytes bytea := '';
> i integer := 0;
> BEGIN
> -- generate 1024 bytes at a time using gen_random_bytes(1024)
> WHILE i < length LOOP
> bytes := bytes || gen_random_bytes(least(1024, length - i));
> i := i + 1024;
> END LOOP;
>
> RETURN bytes;
> END;
> $$ LANGUAGE plpgsql;
>
> DROP TABLE IF EXISTS original, original_hashed, dictionary_table CASCADE;
> DROP VIEW IF EXISTS joined_view;
>
> CREATE TABLE original AS (
> WITH large_random_bytes AS (
> SELECT n AS id, random_bytes(4096 + n) AS attributes
> FROM generate_series(1, 1000) n
> )
> SELECT
> -- An incrementing timestamp
> '2024-01-01 00:00:00.00+00'::timestamptz + (n || ' seconds')::interval
> AS start_timestamp,
> -- Another relatively small column, let's just make it a random string
> including `n` to make it unique
> 'random_string_' || (n % 100) AS event_name,
> -- The attributes column
> lrb.attributes AS attributes
> FROM generate_series(0, 100) n
> JOIN large_random_bytes lrb ON n % 1000 = lrb.id
> );
> CREATE INDEX ON original(start_timestamp);
>
> CREATE TABLE dictionary_table AS (
> SELECT DISTINCT time_bucket('1 day', start_timestamp) AS
> start_timestamp_range, attributes, md5(attributes) AS hash
> FROM original
> );
> CREATE INDEX ON dictionary_table (start_timestamp_range, hash);
>
> CREATE TABLE original_hashed AS (
> SELECT
> start_timestamp,
> event_name,
> md5(attributes) AS hash
> FROM original
> );
> CREATE INDEX ON original_hashed (start_timestamp, time_bucket('1 day',
> start_timestamp), hash);
>
>
> CREATE VIEW joined_view AS (
> SELECT
> original_hashed.start_timestamp,
> original_hashed.event_name,
> dictionary_table.attributes
> FROM original_hashed
> LEFT JOIN dictionary_table ON (
> time_bucket('1 day', original_hashed.start_timestamp) =
> dictionary_table.start_timestamp_range
> AND
> original_hashed.hash = dictionary_table.hash
> )
> );
>
> -- Select all data
> explain (analyze, buffers, verbose, settings, format json)
> SELECT *
> FROM joined_view
> WHERE (
> start_timestamp > ('2024-01-01 00:00:00.00+00'::timestamptz + interval
> '1 day')
> AND
> start_timestamp < ('2024-01-01 00:00:00.00+00'::timestamptz + interval
> '2 day')
> );
> --
> https://app.pgmustard.com/#/explore/f8b3b7d5-c236-4326-9a67-20fa4cdff4bc
>
> explain (analyze, buffers, verbose, settings, format json)
> SELECT *
> FROM original
> WHERE (
> start_timestamp > ('2024-01-01 00:00:00.00+00'::timestamptz + interval
> '1 day')
> AND
> start_timestamp < ('2024-01-01 00:00:00.00+00'::timestamptz + interval
> '2 day')
> );
> --
> https://app.pgmustard.com/#/explore/bcad3767-b8aa-4a37-acc7-7667266f29be
>
> -- A relatively aggregation selective query
> explain (analyze, buffers, verbose, settings, format json)
> SELECT count(*)
> FROM joined_view
> WHERE (
> get_byte(attributes, 4) < 100
> AND
> start_timestamp > ('2024-01-01 00:00:00.00+00'::timestamptz + interval
> '1 day')
> AND
> start_timestamp < ('2024-01-01 00:00:00.00+00'::timestamptz + interval
> '2 day')
> );
> --
> https://app.pgmustard.com/#/explore/c5b68839-4c66-4b14-8de5-757bfc22fdb3
>
> explain (analyze, buffers, verbose, settings, format json)
> SELECT count(*)
> FROM original
> WHERE (
> get_byte(attributes, 4) < 100
> AND
> start_timestamp > ('2024-01-01 00:00:00.00+00'::timestamptz + interval
> '1 day')
> AND
> start_timestamp < ('2024-01-01 00:00:00.00+00'::timestamptz + interval
> '2 day')
> );
> --
> https://app.pgmustard.com/#/explore/831d9c4c-e911-4caf-a6c3-eee5343e44d8
>
> -- Select only some columns
> explain (analyze, buffers, verbose, settings, format json)
> SELECT start_timestamp
> FROM joined_view
> WHERE (
> start_timestamp > ('2024-01-01 00:00:00.00+00'::timestamptz + interval
> '1 day')
> AND
> start_timestamp < ('2024-01-01 00:00:00.00+00'::timestamptz + interval
> '2 day')
> );
> --
> https://app.pgmustard.com/#/explore/5f617bfa-7816-4f8d-9fd7-24388c08978c
>
> explain (analyze, buffers, verbose, settings, format json)
> SELECT start_timestamp
> FROM original
> WHERE (
> start_timestamp > ('2024-01-01 00:00:00.00+00'::timestamptz + interval
> '1 day')
> AND
> start_timestamp < ('2024-01-01 00:00:00.00+00'::timestamptz + interval
> '2 day')
> );
> --
> https://app.pgmustard.com/#/explore/f05de3e8-ca8d-4404-ad33-69f103a3b85a
>
>
> -- A relatively selective query
> explain (analyze, buffers, verbose, settings, format j
Re: Compressing large column by moving it to a unique table
I'm using PostgreSQL 15.5.
Here's a self-contained example. I included links to public pgMustard query
plans.
Gist link: https://gist.github.com/adriangb/05a01cca99a438de2a286945903c44f4
Also copied below for archiving:
```sql
CREATE OR REPLACE FUNCTION random_bytes(length integer)
RETURNS bytea AS $$
DECLARE
bytes bytea := '';
i integer := 0;
BEGIN
-- generate 1024 bytes at a time using gen_random_bytes(1024)
WHILE i < length LOOP
bytes := bytes || gen_random_bytes(least(1024, length - i));
i := i + 1024;
END LOOP;
RETURN bytes;
END;
$$ LANGUAGE plpgsql;
DROP TABLE IF EXISTS original, original_hashed, dictionary_table CASCADE;
DROP VIEW IF EXISTS joined_view;
CREATE TABLE original AS (
WITH large_random_bytes AS (
SELECT n AS id, random_bytes(4096 + n) AS attributes
FROM generate_series(1, 1000) n
)
SELECT
-- An incrementing timestamp
'2024-01-01 00:00:00.00+00'::timestamptz + (n || ' seconds')::interval
AS start_timestamp,
-- Another relatively small column, let's just make it a random string
including `n` to make it unique
'random_string_' || (n % 100) AS event_name,
-- The attributes column
lrb.attributes AS attributes
FROM generate_series(0, 100) n
JOIN large_random_bytes lrb ON n % 1000 = lrb.id
);
CREATE INDEX ON original(start_timestamp);
CREATE TABLE dictionary_table AS (
SELECT DISTINCT time_bucket('1 day', start_timestamp) AS
start_timestamp_range, attributes, md5(attributes) AS hash
FROM original
);
CREATE INDEX ON dictionary_table (start_timestamp_range, hash);
CREATE TABLE original_hashed AS (
SELECT
start_timestamp,
event_name,
md5(attributes) AS hash
FROM original
);
CREATE INDEX ON original_hashed (start_timestamp, time_bucket('1 day',
start_timestamp), hash);
CREATE VIEW joined_view AS (
SELECT
original_hashed.start_timestamp,
original_hashed.event_name,
dictionary_table.attributes
FROM original_hashed
LEFT JOIN dictionary_table ON (
time_bucket('1 day', original_hashed.start_timestamp) =
dictionary_table.start_timestamp_range
AND
original_hashed.hash = dictionary_table.hash
)
);
-- Select all data
explain (analyze, buffers, verbose, settings, format json)
SELECT *
FROM joined_view
WHERE (
start_timestamp > ('2024-01-01 00:00:00.00+00'::timestamptz + interval '1
day')
AND
start_timestamp < ('2024-01-01 00:00:00.00+00'::timestamptz + interval '2
day')
);
-- https://app.pgmustard.com/#/explore/f8b3b7d5-c236-4326-9a67-20fa4cdff4bc
explain (analyze, buffers, verbose, settings, format json)
SELECT *
FROM original
WHERE (
start_timestamp > ('2024-01-01 00:00:00.00+00'::timestamptz + interval '1
day')
AND
start_timestamp < ('2024-01-01 00:00:00.00+00'::timestamptz + interval '2
day')
);
-- https://app.pgmustard.com/#/explore/bcad3767-b8aa-4a37-acc7-7667266f29be
-- A relatively aggregation selective query
explain (analyze, buffers, verbose, settings, format json)
SELECT count(*)
FROM joined_view
WHERE (
get_byte(attributes, 4) < 100
AND
start_timestamp > ('2024-01-01 00:00:00.00+00'::timestamptz + interval '1
day')
AND
start_timestamp < ('2024-01-01 00:00:00.00+00'::timestamptz + interval '2
day')
);
-- https://app.pgmustard.com/#/explore/c5b68839-4c66-4b14-8de5-757bfc22fdb3
explain (analyze, buffers, verbose, settings, format json)
SELECT count(*)
FROM original
WHERE (
get_byte(attributes, 4) < 100
AND
start_timestamp > ('2024-01-01 00:00:00.00+00'::timestamptz + interval '1
day')
AND
start_timestamp < ('2024-01-01 00:00:00.00+00'::timestamptz + interval '2
day')
);
-- https://app.pgmustard.com/#/explore/831d9c4c-e911-4caf-a6c3-eee5343e44d8
-- Select only some columns
explain (analyze, buffers, verbose, settings, format json)
SELECT start_timestamp
FROM joined_view
WHERE (
start_timestamp > ('2024-01-01 00:00:00.00+00'::timestamptz + interval '1
day')
AND
start_timestamp < ('2024-01-01 00:00:00.00+00'::timestamptz + interval '2
day')
);
-- https://app.pgmustard.com/#/explore/5f617bfa-7816-4f8d-9fd7-24388c08978c
explain (analyze, buffers, verbose, settings, format json)
SELECT start_timestamp
FROM original
WHERE (
start_timestamp > ('2024-01-01 00:00:00.00+00'::timestamptz + interval '1
day')
AND
start_timestamp < ('2024-01-01 00:00:00.00+00'::timestamptz + interval '2
day')
);
-- https://app.pgmustard.com/#/explore/f05de3e8-ca8d-4404-ad33-69f103a3b85a
-- A relatively selective query
explain (analyze, buffers, verbose, settings, format json)
SELECT *
FROM joined_view
WHERE (
get_byte(attributes, 4) < 100
AND
start_timestamp > ('2024-01-01 00:00:00.00+00'::timestamptz + interval '1
day')
AND
start_timestamp < ('2024-01-01 00:00:00.00+00'::timestamptz + interval '2
day')
);
-- https://app.pgmustard.com/#/explore/a75b20bd-07d5-4402-a0d8-22419682307a
explain (analyze, buffers, verbose, settings, format json)
SELECT *
FROM original
WHERE (
get_byte(attributes, 4) < 100
AND
start_timestamp > ('2024-01-01
Re: Compressing large column by moving it to a unique table
1. Show us the PG version, view definition, the exact query that's slow,
and the EXPLAIN (ANALYZE).
2. Presumably there's an index on each table's *_hash column?
On Tue, Feb 13, 2024 at 8:48 AM Adrian Garcia Badaracco
wrote:
> Thank you for the reply Ron.
>
> Yes there are many fewer (<1%) the number of rows in new_table.
>
> Thanks for making me think of normalization, I hadn’t seen it that way.
> Although there is no theoretical relationship between the rows in the other
> columns in the original table and the attributes column, in practice there
> is a strong correlation, so I guess what I am trying to capture here is
> taking advantage of that correlation, while not completely depending on it
> because it can be broken.
>
> In any case, whatever theoretical framework is put around this solution, I
> am also interested in the practical aspects, in particular that case of
> selecting a subset of columns from the view that I know doesn’t need the
> join but the query planner thinks does.
>
> On Tue, Feb 13, 2024 at 3:16 AM Ron Johnson
> wrote:
>
>> On Mon, Feb 12, 2024 at 10:12 PM Adrian Garcia Badaracco <
>> adr...@adriangb.com> wrote:
>>
>>> I am using Timescale so I'll be mentioning some timestamp stuff but I
>>> think this is a general postgres question for the most part.
>>>
>>> I have a table with some fixed, small columns (id, timestamp, etc) and a
>>> large JSONB column (let's call it `attributes`). `attributes` has 1000s of
>>> schemas, but given a schema, there's a lot of duplication. Across all rows,
>>> more than 99% of the data is duplicated (as measured by `count(attributes)`
>>> vs `count(distinct attributes)`.
>>>
>>> I can't normalize `attributes` into real columns because it is quite
>>> variable (remember 1000s of schemas).
>>>
>>> My best idea is to make a table like `(day timestamptz, hash text,
>>> attributes jsonb)` and then in my original table replace `attributes` with
>>> a reference to `new_table`.
>>>
>>
>> Meaning that there are many fewer rows in new_table?
>>
>>
>>> I can then make a view that joins them `select original_table.timestamp,
>>> new_table.attributes from original join new_table on (time_bucket('1 day',
>>> timestamp) = day AND original.hash = new_table.hash)` or something like
>>> that. The idea of time bucketing into 1 day is to balance write and read
>>> speed (by relying on timescale to do efficient time partitioning, data
>>> retention, etc.).
>>>
>>
>>> I recognize this is essentially creating a key-value store in postgres
>>> and also janky compression, so I am cautious about it.
>>>
>>
>> If my interpretation (that there are many fewer rows in new_table) is
>> correct, then you've stumbled into the Second Normal Form of database
>> design: https://en.wikipedia.org/wiki/Second_normal_form#Example
>>
>>
Re: Compressing large column by moving it to a unique table
Thank you for the reply Ron.
Yes there are many fewer (<1%) the number of rows in new_table.
Thanks for making me think of normalization, I hadn’t seen it that way.
Although there is no theoretical relationship between the rows in the other
columns in the original table and the attributes column, in practice there
is a strong correlation, so I guess what I am trying to capture here is
taking advantage of that correlation, while not completely depending on it
because it can be broken.
In any case, whatever theoretical framework is put around this solution, I
am also interested in the practical aspects, in particular that case of
selecting a subset of columns from the view that I know doesn’t need the
join but the query planner thinks does.
On Tue, Feb 13, 2024 at 3:16 AM Ron Johnson wrote:
> On Mon, Feb 12, 2024 at 10:12 PM Adrian Garcia Badaracco <
> adr...@adriangb.com> wrote:
>
>> I am using Timescale so I'll be mentioning some timestamp stuff but I
>> think this is a general postgres question for the most part.
>>
>> I have a table with some fixed, small columns (id, timestamp, etc) and a
>> large JSONB column (let's call it `attributes`). `attributes` has 1000s of
>> schemas, but given a schema, there's a lot of duplication. Across all rows,
>> more than 99% of the data is duplicated (as measured by `count(attributes)`
>> vs `count(distinct attributes)`.
>>
>> I can't normalize `attributes` into real columns because it is quite
>> variable (remember 1000s of schemas).
>>
>> My best idea is to make a table like `(day timestamptz, hash text,
>> attributes jsonb)` and then in my original table replace `attributes` with
>> a reference to `new_table`.
>>
>
> Meaning that there are many fewer rows in new_table?
>
>
>> I can then make a view that joins them `select original_table.timestamp,
>> new_table.attributes from original join new_table on (time_bucket('1 day',
>> timestamp) = day AND original.hash = new_table.hash)` or something like
>> that. The idea of time bucketing into 1 day is to balance write and read
>> speed (by relying on timescale to do efficient time partitioning, data
>> retention, etc.).
>>
>
>> I recognize this is essentially creating a key-value store in postgres
>> and also janky compression, so I am cautious about it.
>>
>
> If my interpretation (that there are many fewer rows in new_table) is
> correct, then you've stumbled into the Second Normal Form of database
> design: https://en.wikipedia.org/wiki/Second_normal_form#Example
>
>
Re: Compressing large column by moving it to a unique table
On Mon, Feb 12, 2024 at 10:12 PM Adrian Garcia Badaracco <
adr...@adriangb.com> wrote:
> I am using Timescale so I'll be mentioning some timestamp stuff but I
> think this is a general postgres question for the most part.
>
> I have a table with some fixed, small columns (id, timestamp, etc) and a
> large JSONB column (let's call it `attributes`). `attributes` has 1000s of
> schemas, but given a schema, there's a lot of duplication. Across all rows,
> more than 99% of the data is duplicated (as measured by `count(attributes)`
> vs `count(distinct attributes)`.
>
> I can't normalize `attributes` into real columns because it is quite
> variable (remember 1000s of schemas).
>
> My best idea is to make a table like `(day timestamptz, hash text,
> attributes jsonb)` and then in my original table replace `attributes` with
> a reference to `new_table`.
>
Meaning that there are many fewer rows in new_table?
> I can then make a view that joins them `select original_table.timestamp,
> new_table.attributes from original join new_table on (time_bucket('1 day',
> timestamp) = day AND original.hash = new_table.hash)` or something like
> that. The idea of time bucketing into 1 day is to balance write and read
> speed (by relying on timescale to do efficient time partitioning, data
> retention, etc.).
>
> I recognize this is essentially creating a key-value store in postgres and
> also janky compression, so I am cautious about it.
>
If my interpretation (that there are many fewer rows in new_table) is
correct, then you've stumbled into the Second Normal Form of database
design: https://en.wikipedia.org/wiki/Second_normal_form#Example
