Skip to main content

Full-text Search using Inverted Indexes [experimental]

Inverted indexes are an experimental type of secondary indexes which provide fast text search capabilities for String or FixedString columns. The main idea of an inverted index is to store a mapping from "terms" to the rows which contain these terms. "Terms" are tokenized cells of the string column. For example, the string cell "I will be a little late" is by default tokenized into six terms "I", "will", "be", "a", "little" and "late". Another kind of tokenizer is n-grams. For example, the result of 3-gram tokenization will be 21 terms "I w", " wi", "wil", "ill", "ll ", "l b", " be" etc. The more fine-granular the input strings are tokenized, the bigger but also the more useful the resulting inverted index will be.

note

Inverted indexes are experimental and should not be used in production environments yet. They may change in the future in backward-incompatible ways, for example with respect to their DDL/DQL syntax or performance/compression characteristics.

Usage​

To use inverted indexes, first enable them in the configuration:

SET allow_experimental_inverted_index = true;

An inverted index can be defined on a string column using the following syntax

CREATE TABLE tab
(
    `key` UInt64,
    `str` String,
    INDEX inv_idx(str) TYPE inverted(0) GRANULARITY 1
)
ENGINE = MergeTree
ORDER BY key

where N specifies the tokenizer:

  • inverted(0) (or shorter: inverted()) set the tokenizer to "tokens", i.e. split strings along spaces,
  • inverted(N) with N between 2 and 8 sets the tokenizer to "ngrams(N)"

The maximum rows per postings list can be specified as the second parameter. This parameter can be used to control postings list sizes to avoid generating huge postings list files. The following variants exist:

  • inverted(ngrams, max_rows_per_postings_list): Use given max_rows_per_postings_list (assuming it is not 0)
  • inverted(ngrams, 0): No limitation of maximum rows per postings list
  • inverted(ngrams): Use a default maximum rows which is 64K.

Being a type of skipping index, inverted indexes can be dropped or added to a column after table creation:

ALTER TABLE tab DROP INDEX inv_idx;
ALTER TABLE tab ADD INDEX inv_idx(s) TYPE inverted(2);

To use the index, no special functions or syntax are required. Typical string search predicates automatically leverage the index. As examples, consider:

INSERT INTO tab(key, str) values (1, 'Hello World');
SELECT * from tab WHERE str == 'Hello World';
SELECT * from tab WHERE str IN ('Hello', 'World');
SELECT * from tab WHERE str LIKE '%Hello%';
SELECT * from tab WHERE multiSearchAny(str, ['Hello', 'World']);
SELECT * from tab WHERE hasToken(str, 'Hello');

The inverted index also works on columns of type Array(String), Array(FixedString), Map(String) and Map(String).

Like for other secondary indices, each column part has its own inverted index. Furthermore, each inverted index is internally divided into "segments". The existence and size of the segments are generally transparent to users but the segment size determines the memory consumption during index construction (e.g. when two parts are merged). Configuration parameter "max_digestion_size_per_segment" (default: 256 MB) controls the amount of data read consumed from the underlying column before a new segment is created. Incrementing the parameter raises the intermediate memory consumption for index construction but also improves lookup performance since fewer segments need to be checked on average to evaluate a query.

Full-text search of the Hacker News dataset​

Let's look at the performance improvements of inverted indexes on a large dataset with lots of text. We will use 28.7M rows of comments on the popular Hacker News website. Here is the table without an inverted index:

CREATE TABLE hackernews (
    id UInt64,
    deleted UInt8,
    type String,
    author String,
    timestamp DateTime,
    comment String,
    dead UInt8,
    parent UInt64,
    poll UInt64,
    children Array(UInt32),
    url String,
    score UInt32,
    title String,
    parts Array(UInt32),
    descendants UInt32
)
ENGINE = MergeTree
ORDER BY (type, author);

The 28.7M rows are in a Parquet file in S3 - let's insert them into the hackernews table:

INSERT INTO hackernews
    SELECT * FROM s3Cluster(
        'default',
        'https://datasets-documentation.s3.eu-west-3.amazonaws.com/hackernews/hacknernews.parquet',
        'Parquet',
        '
    id UInt64,
    deleted UInt8,
    type String,
    by String,
    time DateTime,
    text String,
    dead UInt8,
    parent UInt64,
    poll UInt64,
    kids Array(UInt32),
    url String,
    score UInt32,
    title String,
    parts Array(UInt32),
    descendants UInt32');

Consider the following simple search for the term ClickHouse (and its varied upper and lower cases) in the comment column:

SELECT count()
FROM hackernews
WHERE hasToken(lower(comment), 'clickhouse');

Notice it takes 3 seconds to execute the query:

┌─count()─┐
│    1145 │
└─────────┘

1 row in set. Elapsed: 3.001 sec. Processed 28.74 million rows, 9.75 GB (9.58 million rows/s., 3.25 GB/s.)

We will use ALTER TABLE and add an inverted index on the lowercase of the comment column, then materialize it (which can take a while - wait for it to materialize):

ALTER TABLE hackernews
     ADD INDEX comment_lowercase(lower(comment)) TYPE inverted;

ALTER TABLE hackernews MATERIALIZE INDEX comment_lowercase;

We run the same query...

SELECT count()
FROM hackernews
WHERE hasToken(lower(comment), 'clickhouse')

...and notice the query executes 4x faster:

┌─count()─┐
│    1145 │
└─────────┘

1 row in set. Elapsed: 0.747 sec. Processed 4.49 million rows, 1.77 GB (6.01 million rows/s., 2.37 GB/s.)

We can also search for one or all of multiple terms, i.e., disjunctions or conjunctions:

-- multiple OR'ed terms
SELECT count(*)
FROM hackernews
WHERE multiSearchAny(lower(comment), ['oltp', 'olap']);

-- multiple AND'ed terms
SELECT count(*)
FROM hackernews
WHERE hasToken(lower(comment), 'avx') AND hasToken(lower(comment), 'sve');
note

Unlike other secondary indices, inverted indexes (for now) map to row numbers (row ids) instead of granule ids. The reason for this design is performance. In practice, users often search for multiple terms at once. For example, filter predicate WHERE s LIKE '%little%' OR s LIKE '%big%' can be evaluated directly using an inverted index by forming the union of the row id lists for terms "little" and "big". This also means that the parameter GRANULARITY supplied to index creation has no meaning (it may be removed from the syntax in the future).

Related Content​