Smart Correlation Repository Notes

Last modified 19 Jul 2022 16:05 +02:00

Table design

One obvious option for identity attributes was to use extension-like JSONB column mechanism. Because there are multiple identity variants for each focus, it would be stored in a separate table with orig and norm variants in different JSONB columns to keep the structure flat.

Levenshtein considerations

Experiments with 10M row table and JSONB column showed expected results when Levenshtein function was used:

  • Full seq-scan must be used, there is no way to index for Levenshtein and various inputs.

  • Depending on the HW, no more than 10s should be expected for such a search on 10M variants. This means the more variants are used per user, the slower the search is. Also, the result may change with the volume (physical size) of the table, because all of it must be fetched.

  • When no indexing is used, there is just small difference whether the JSONB values are singlevalued or multivalued (which requires exists subquery with from jsonb_array_elements_text).

Trigram indexing

For some operations (ilike and similarity) trigram index can be successfully used. This requires pg_trgm extension which is already enabled in our DB schema.

  • Index must be created with this type gin(<column> gin_trgm_ops), where <column> is the actual column name - it can be of text and compatible types - it does not work for arrays and JSONB.

  • JSONB single value can be indexed using gin((attrs →> 'first_name') gin_trgm_ops). With this index operations like (attrs →> 'first_name') % 'brady' are very fast.

However, trigram index can’t be used to array values, in our case JSONB array values. There are some options with arrays:

  • exists subquery with from jsonb_array_elements_text can be used - which does not use any index, of course.

  • Limited for some operations: We can extract/normalize the multivalue into a single value string (similar to fullTextInfo used in MP) and then index this and add where to it as well. Additional conditions can be used as well; e.g. using trigram only for the digest column, this makes the query fast, and then check all the "slow" conditions on the other attributes (even multivalued). This solution is more complicated, because it requires some auto-magic either on repository or more work on the model level.

    • Instead of explicit column we can create index using gin((attrs →> 'multi-val-key') gin_trgm_ops). This indexes textual representation of the whole array value, so it’s not usable as trigram index for each value. This can be used instead of explicit index column, it works just as well although it still allows for false positives - thuese must be filtered out by proper additional conditions. Related fast check then is attrs →> 'multi-value-key' ilike '%value%'. Note that this "fast check" is usable only for some operations, e.g. like works fine, but similarity (%) does not. The reason is that while the single name is similar enough to the provided value, the whole digest returns false on the same check and would falsly ruled valid results out.

    Doing this blindly without index does not help, this would require some additional information that should be stored in m_ext_item catalog in some additional configuration column (which would also be of JSONB type for flexibility, but this is unrelated).

    Perhaps it’s better not to consider this for now. The added complexity is probably not worth the questionable benefits.

In any case, using JSONB seems to be a better way in overall. Using the detail table for key-value pairs takes roughly the same space, but requires join (or semi-join, that is exists) and even with indexes does not perform faster.

It is expected that both single table or master-detail variant will experience some level of update bloat, but it should be managable by default auto-vacuum.

SQL experiments code:

CREATE EXTENSION fuzzystrmatch;

create table x_names (
    id SERIAL NOT NULL PRIMARY KEY,
    first_name text,
    last_name text,
    gender varchar(1),
    country varchar(2),
    attrs jsonb
);

create table x_names_attrs (
    id SERIAL NOT NULL PRIMARY KEY,
    name_id integer NOT NULL,
    key text,
    value text
);

ALTER TABLE x_names_attrs ADD CONSTRAINT x_names_attrs_name_id_fk
    FOREIGN KEY (name_id) REFERENCES x_names (id)
        ON DELETE CASCADE;

CREATE INDEX x_names_attrs_name_id_idx ON x_names_attrs (name_id);
-- alternative to pure FK index, does not seem to help with the queries we need
CREATE INDEX x_names_attrs_name_id_key_idx ON x_names_attrs (name_id, key);

-- copy of names from CSV file, in the end 10M names were used for experiments
COPY x_names(first_name, last_name, gender, country)
FROM '/vagrant/us.csv'
DELIMITER ','
;

-- lateral join, seems fine when we need l and m in select too, but is slower than the next select
SELECT count(*)
from x_names,
     levenshtein_less_equal(first_name, 'Brad', 5) l,
     levenshtein_less_equal(last_name, 'Brad', 5) m
where l <= 5 or m <= 5
;

-- fastest variant, using columns
SELECT count(*)
from x_names
where levenshtein_less_equal(first_name, 'Brad', 5) <= 5
      or levenshtein_less_equal(last_name, 'Brad', 5) <= 5

-- filling the JSONB column, choose small or larger JSON content, it changes the speeds of full-scan
update x_names
-- small JSONB
set attrs = ('{"first_name": "' || first_name || '", "last_name": "' || last_name || '", "names": ["' || first_name || '", "' || last_name || '"]}')::jsonb
-- bit bigger JSONB
-- set attrs = ('{"first_name": "' || first_name || '", "last_name": "' || last_name || '", "names": ["' || first_name || '", "' || last_name || '"],' ||
-- ' "xxx1": "yyy1", "xxx2": "yyy2", "xxx3": "yyy3", "xxx4": "yyy4", "xxx5": "yyy5", "xxx6": "yyy7", "xxx7": "yyy7", "xxx8": "yyy8",' ||
-- ' "xxx11": "yyy11", "xxx12": "yyy12", "xxx13": "yyy13", "xxx14": "yyy14", "xxx15": "yyy15", "xxx16": "yyy16", "xxx17": "yyy17", "xxx18": "yyy18",' ||
-- ' "xxx21": "yyy21", "xxx22": "yyy22", "xxx23": "yyy23", "xxx24": "yyy24", "xxx25": "yyy25", "xxx26": "yyy26", "xxx27": "yyy27", "xxx28": "yyy28"}')::jsonb
;

-- Showing matching names (from first 50), unrolling the names multi-val JSON value.
-- This, again, uses lateral join, which is not best for WHERE, see the query later.
select *
from x_names,
     jsonb_array_elements_text(attrs -> 'names') n,
     levenshtein_less_equal(n, 'Brad', 5) l
where id <= 50
      and l <  5
;

-- Using single value from JSONB (similar to single column, just marginally slower).
SELECT count(*)
from x_names
where levenshtein_less_equal(attrs ->> 'first_name', 'Brad', 5) <= 5;

-- This checks any of the multi-value names:
SELECT count(*)
from x_names
where exists (select 1 from jsonb_array_elements_text(attrs -> 'names') n
    where levenshtein_less_equal(n, 'Brad', 5) <= 5)

-- Now filling the detail table:
insert into x_names_attrs (name_id, key, value)
select id, key, coalesce(aval, sval) val from (
    select id, key,
        case when jsonb_typeof(value) = 'array' then value end avals,
        case when jsonb_typeof(value) <> 'array' then value end sval
    from x_names, jsonb_each(attrs)
    where id > 1000000
) x left join jsonb_array_elements(avals) aval on true;

-- query using detail table and two "columns"
SELECT count(*)
from x_names n
where exists(select 1 from x_names_attrs
    where name_id = n.id
        and (key = 'first_name' and levenshtein_less_equal(value, 'Brad', 5) <= 5
            or key = 'last_name' and levenshtein_less_equal(value, 'Brad', 5) <= 5));

-- detail table, query for multi-val names key
SELECT count(*)
from x_names
where exists(select 1 from x_names_attrs
    where name_id = n.id
        and key = 'names'
        and levenshtein_less_equal(value, 'Brad', 5) <= 5);

-- For other operations like % and LIKE/ILIKE we can build trigram indexes:
-- One for specific key in JSON (all values must be single-value texts):
create index x_names_afn_trg_idx
    on x_names using gin((attrs ->> 'first_name') gin_trgm_ops);

-- Tried this to make the query for 'names' in detail table faster, but did not work.
create index x_names_attrs_names_trg_idx
    on x_names_attrs using gin(value gin_trgm_ops)
    where key = 'names';

-- Very fast query using JSONB and trigram index (note % operation, levenshtein does not use index):
SELECT count(*)
from x_names n
where (attrs ->> 'first_name') % 'brady';

Talking about speeds of the queries in very broad terms (can vary on HW, PG config…​):

  • JSONB single value silimarity (%) search using index is very fast (under 1s counting 17k matches from 10M rows).

  • Levenshtein check on column or JSONB single value is ~10s.

  • Levenshtein check on multi-val JSONB (with subquery unrolling values) is ~20s.

  • Levenshtein check on multi-val names in detail table is ~50s.

  • Similarity search on multi-value JSONB (with subquery unrolling values) is ~30s.

  • Similarity search on multi-value names using detail table is ~30s. This is 20s less than the same with levenshtein function which can’t be attributed only to the levenshtein performance (the same number of calls is performed for 10s query too). The similarity query uses Bitmap Index Scan on x_names_attrs_names_trg_idx which helps a bit.

  • Any lateral join makes the 10s query go up above 30s, even for 1:1 joins on levenshtein functions. We will not use these variants, they may be practical only for getting some values into select.

Custom operations and Querydsl

Was this page helpful?
YES NO
Thanks for your feedback