Analytics/Data Lake

From Wikitech
Jump to navigation Jump to search

The Analytics Data Lake (ADL), or the Data Lake for short, is a large, analytics-oriented repository of data about Wikimedia projects (in industry terms, a data lake).

Data available

Currently, you need production data access to use some of this data. A lot of it is available publicly at dumps.wikimedia.org.
Traffic data
Webrequest, pageviews, and unique devices
Edits data
Historical data about revisions, pages, and users (e.g. MediaWiki History)
Content data
Wikitext (latest & historical) and wikidata-entities
Events data
EventLogging, EventBus and event streams data (raw, refined, sanitized)

Some of these datasets (such as webrequests) are only available in Hive, while others (such as pageviews) are also available as data cubes (usually in more aggregated capacity).

Access

The main way to access the data in the Data Lake is to run queries using one of the three available SQL engines: Presto, Hive, and Spark.

You can access these engines through several different routes:

All three engines also have command-line programs which you can use on one of the analytics clients. This is probably the least convenient way, but if you want to use it, consult the engine's documentation page.

Syntax differences between the SQL engines

For the most part, Presto, Hive, and Spark work the same way, but they have some differences in SQL syntax.

use case Spark Presto Hive
keyword for the string data type STRING VARCHAR STRING
string literal 'foo', "foo" 'foo' 'foo', "foo"
keyword for 32-bit float data type FLOAT, REAL REAL FLOAT
keyword for 64-bit float data type DOUBLE
select a column named with a reserved word (e.g. DATE) `date` "date" `date`
get the length of an array SIZE(a) CARDINALITY(a) SIZE(a)
concatenate strings with a separator CONCAT_WS not available CONCAT_WS
count rows which match a condition COUNT_IF(x = y) COUNT_IF(x = y) SUM(CAST(x = y AS INT))
transform integer year/month/day fields to a date string CONCAT(year, '-', LPAD(month, 2, '0'), '-', LPAD(day, 2, '0')) CONCAT(CAST(year AS VARCHAR), '-', LPAD(CAST(month AS VARCHAR), 2, '0'), '-', LPAD(CAST(day AS VARCHAR), 2, '0')) CONCAT(year, '-', LPAD(month, 2, '0'), '-', LPAD(day, 2, '0'))
convert an ISO 8601 timestamp string (e.g. "2021-11-01T01:23:02Z") to an SQL timestamp TO_TIMESTAMP(dt) FROM_ISO8601_TIMESTAMP(dt) FROM_UNIXTIME(UNIX_TIMESTAMP(dt, "yyyy-MM-dd'T'HH:mm:ss'Z'"))
divide integers, returning a float if necessary x / y CAST(x AS DOUBLE) / y x / y
select the first or last rows in a group FIRST, LAST not available not available
  • It's useful to get in the habit of using singled quoted text ('foo') for strings, since all three engines interpret it the same way. Double quoted text ("foo") is interpreted as a string in Spark and Hive, but as a column name in Presto.
  • Escaping special characters in string literals works differently in Spark and Presto. See this notebook for more details.
  • See also: Presto's guide to migrating from Hive

Integer division in Presto

If you divide integers, Hive and Spark will return a floating-point number if necessary (e.g. 1 / 3 returns 0.333333). However, Presto will return only an integer (e.g. 1 / 3 returns 0). Use CAST(x AS DOUBLE) to work around this. DOUBLE is a 64-bit floating point number, while REAL is a 32-bit floating point number.

There are some quirks to be aware of with this behavior:

SELECT
  2/5 AS "none",
  CAST(2 AS DOUBLE)/5 AS "numerator",
  2/CAST(5 AS DOUBLE) AS "denominator",
  CAST(2/5 AS DOUBLE) AS "outer",
  2/5 * CAST(100 AS DOUBLE) AS "percentage (a)",
  CAST(2/5 AS DOUBLE) * 100 AS "percentage (b)",
  CAST(2 AS DOUBLE) / 5 * 100 AS "percentage (c)",
  1.0 * 2 / 5 AS "percentage (d)"

These produce:

  • none: 0 (because 2/5 is rounded towards 0 to keep the output data type integer, same as input)
  • numerator, denominator: 0.4
  • outer: 0 (because 2/5 is implicitly cast to integer BEFORE being explicitly cast as double)
  • percentage
    • (a): 0 (same as "none" – 2/5 is cast to int and rounded towards 0 before it reaches the double-typed 100)
    • (b): 0 (same as outer)
    • (c): 40
    • (d): 40

So let's say your query has SUM(IF(event.action = 'click', 1, 0)) / COUNT(1) to calculate clickthrough rate. It'll be 0 unless you:

  • explicitly cast either the denominator or the numerator to double, or
  • implicitly cast by multiplying by 1.0 (for example above it follows order of operations: 1.0 * 2 becomes 2.0 then that gets divided by 5)

Table and file formats

Data Lake tables can be created using either Hive format or Iceberg format. Iceberg is the successor to Hive, and highly recommended for new tables. As of Feb 2024, the existing tables in the wmf database are being slowly migrated to Iceberg (task T333013).

Both table formats can store data using a variety of underlying file formats; we normally use Parquet with both Hive and Iceberg.

Technical architecture

Data Lake datasets which are available in Hive are stored in the Hadoop Distributed File System (HDFS). The Hive metastore is a centralized repository for metadata about these data files, and all three SQL query engines we use (Presto, Spark SQL, and Hive) rely on it.

Some Data Lake datasets are available in Druid, which is separate from Hive and HDFS, and allows quick exploration and dashboarding of those datasets in Turnilo and Superset.

The Analytics cluster, which consists of Hadoop servers and related components, provides the infrastructure for the Data Lake.

All Subpages of Analytics/Data Lake