CREATE SOURCE: Text or bytes from an S3 bucket

CREATE SOURCE connects Materialize to an external data source and lets you interact with its data as if the data were in a SQL table.

This document details how to connect Materialize to an S3 Bucket that contains multiple objects, and to listen for new object creation. Each S3 object can contain multiple byte- or text-encoded records, separated by newlines.

Conceptual framework

Sources represent connections to resources outside Materialize that it can read data from. For more information, see API Components: Sources.

Syntax

key_constraint

with_options

WITH options

The following options are valid within the WITH clause.

AWS Credentials WITH options

If you do not provide credentials via with options then materialized will examine the standard AWS authorization chain:

1. Environment variables: AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY
2. credential_process command in the AWS config file, usually located at ~/.aws/config.
3. AWS credentials file. Usually located at ~/.aws/credentials.
4. IAM instance profile. Will only work if running on an EC2 instance with an instance profile/role.

Credentials fetched from a container or instance profile expire on a fixed schedule. Materialize will attempt to refresh the credentials automatically before they expire, but the source will become inoperable if the refresh operation fails.

Permissions Required

The IAM User or Role used by materialized requires permission to perform different AWS actions depending on which actions are required to discover the list of objects to download.

The DISCOVER OBJECTS USING clause describes how Materialize will load objects, and so its parameters determine which permissions materialized requires. For example, since the SCAN key name source (as in DISCOVER OBJECTS USING BUCKET SCAN) must perform repeated ListObjects actions to create a list of key names to download, you must grant the Materialize IAM User or Role the ListObjects permission before you specify DISCOVER OBJECTS USING BUCKET SCAN.

The root AWS documentation for S3 permissions is available here.

Details

S3 source details

The S3 source is designed to ingest a large volume of static data from AWS’s Simple Storage Service, Amazon Web Services’ cloud object store. It is important to note that an object store behaves differently than a file system does, and that those differences affect how Materialize interacts with objects in it.

Some key differences from file systems include:

• Latency on individual S3 operations is much higher than on even cloud-network filesystems.
• It is not possible to seek to a point in an object; Materialize must download the entire object from the beginning to read to a point.
• All object operations are atomic. It is not possible to modify just part of an S3 object; you can only entirely replace or delete objects.

The primary effect of this is that we do not handle updates to S3 objects, and we may interleave multiple object ingestion to speed it up.

S3 limitations

• Currently S3 sources do not support Avro- or Protobuf- encoded objects. Implementation issues: Avro, Protobuf
• Object ingest order is not guaranteed.
• All S3 sources are append-only. Deleted and updated S3 objects are silently ignored.

Object discovery strategies

Materialize has several techniques to discover the objects to download, with more planned for the future.

All strategies follow the same basic pattern. Materialize will:

• Obtain a list of objects using AWS APIs.
• Deduplicate objects so the same object is never downloaded twice.
• Filter the list of objects against the MATCHING clause patterns.
• Download the matching objects.
• Treat each object downloaded as a newline-delimited file for the purposes of record delineation.

No guarantees are made about the relative order of records sent from the object discovery process and the Materialize SQL engine. As usual with SQL, you must impose your own order.

You may specify strategies multiple times within one CREATE SOURCE statement. For example, this is a legal invocation:

DISCOVER OBJECTS USING
  BUCKET SCAN 'example',
  BUCKET SCAN 'other',
  SQS NOTIFICATIONS 'example-notifications'

Listing Bucket objects

The BUCKET SCAN discovery performs a single scan over the specified bucket at source creation time. If you would like an S3 source to ingest objects that are added to the bucket after the source is created you must also configure an SQS NOTIFICATIONS discovery mechanism on the source.

Listening to SQS notifications

AWS S3 has a built-in method for notifying downstream applications of bucket modification, the Event Notification API. For Materialize, the only interesting modifications are object creation, aka the s3:ObjectCreated:* event namespace. Follow the AWS Tutorial to configure a bucket for exactly this namespace.

Once you have configured S3 notifications to go to an SQS queue, you can point Materialize at that queue with DISCOVER OBJECTS USING SQS NOTIFICATIONS 'queue-name'.

Materialize deletes SQS messages as soon as they are ingested. This means that the same SQS queue cannot be used for multiple sources. If you would like to have multiple sources listening to notifications from the same bucket you must configure an SNS topic as an intermediary, with multiple SQS queues subscribed to it, one SQS queue per source. Note that SQS queues subscribed to SNS topics intended for Materialize must be configured to use raw message delivery.

Since Materialize treats unmaterialized sources with multiple downstream views as separate sources, SQS NOTIFICATIONS should not be used with unmaterialized sources. This behavior of unmaterialized sources is considered a bug (#7423), and will be improved in a future release.

Patterns

It is possible to filter the list of object keys to download using unix-style glob syntax as an argument to the MATCHING clause:

• ? matches any single character except /.
• * matches zero or more characters except /.
• ** recursively matches directories, but some other pattern must be specified. For example, a/** matches anything inside of the a/ prefix (but not a/ itself), and **/a matches a in any prefix, but not a with no prefix.
• {a,b} matches a or b where a and b are arbitrary glob patterns.
• [ab] matches a or b where a and b are characters. Prepend ! to the matched characters to invert the match, e.g. [!ab] matches any character besides a or b.
• You can escape metacharacters such as * and ? with character class notation. For example, [*] matches *.

Pattern examples

Compression

• Omitting COMPRESSION is equivalent to COMPRESSION NONE.
• Using GZIP compression requires the object be compressed using the gzip algorithm or that it be a concatenation of multiple gzip member streams.

Regex format details

Regex-formatted sources let you apply a structure to arbitrary strings passed in from file sources. This is particularly useful when processing unstructured log files.

• To parse regex strings, Materialize uses rust-lang/regex. For more detail, refer to its documented syntax.
• To create a column in the source, create a capture group, i.e. a parenthesized expression, e.g. ([0-9a-f]{8}).
  • Name columns by…

    • Using the col_name option when creating the source. The number of names provided must match the number of capture groups.

    • Creating named captured groups, e.g. ?P<offset> in (?P<offset>[0-9a-f]{8}) creates a column named offset.

      Unnamed capture groups are named column1, column2, etc.

• We discard all data not included in a capture group. You can create non-capturing groups using ?: as the leading pattern in the group, e.g. (?:[0-9a-f]{4} ){8}.

Text format details

Text-formatted sources reads lines from a file.

• Data from text-formatted sources is treated as newline-delimited.
• Data is assumed to be UTF-8 encoded, and discarded if it cannot be converted to UTF-8.
• Text-formatted sources have one column, which, by default, is named text.

Raw byte format details

Raw byte-formatted sources provide Materialize the raw bytes received from the source without applying any formatting or decoding.

Raw byte-formatted sources have one column, which, by default, is named data.

Append-only envelope

Append-only envelope means that all records received by the source is treated as an insert. This is Materialize’s default envelope (i.e. if no envelope is specified), and can be specified with ENVELOPE NONE.

Key constraint details

Primary keys are automatically inferred for Kafka sources using Upsert or Debezium envelopes and Postgres sources.

For other source configurations, the key_constraint syntax allows to manually declare a set of columns as a primary key. This enables optimizations and constructs that rely on a key to be present when it cannot be inferred.

WARNING! Materialize will not enforce the constraint and will produce wrong results if it is not correct.

Examples

Assuming there is an S3 bucket named “frontend” that contains the following keys and associated content:

logs/2020/12/31/frontend.log

99.99.44.44 - - [12/31/2020:23:55:59 +0000] "GET /updates HTTP/1.1" 200 10020 "-" "Python/Requests 22"
80.91.33.133 - - [12/31/2020:23:55:02 +0000] "GET /downloads/materialized HTTP/1.1" 304 0 "-" "Rust/reqwest 0.3"
173.203.139.108 - - [12/31/2020:23:55:07 +0000] "GET /wpadmin HTTP/1.1" 404 332 "-" "Firefox 9000"
173.203.139.108 - - [12/31/2020:23:55:14 +0000] "GET /downloads/materialized HTTP/1.1" 404 334 "-" "Python/Requests 22"
99.99.44.44 - - [12/31/2020:23:55:01 +0000] "GET /downloads/materialized HTTP/1.1" 304 0 "-" "Python/Requests 22"
80.91.33.133 - - [12/31/2020:23:55:41 +0000] "GET /downloads/materialized HTTP/1.1" 304 0 "-" "Rust/reqwest 0.3"
37.26.93.214 - - [12/31/2020:23:55:52 +0000] "GET /updates HTTP/1.1" 200 3318 "-" "Go 1.1 package http"

logs/2021/01/01/frontend.log

99.99.44.44 - - [01/01/2021:00:00:41 +0000] "GET /updates HTTP/1.1" 200 10020 "-" "Python/Requests 22"
188.138.60.101 - - [01/01/2021:00:00:48 +0000] "GET /updates HTTP/1.1" 200 10020 "-" "Python/Requests 22"
46.50.21.83 - - [01/01/2021:00:00:02 +0000] "GET /downloads/materialized HTTP/1.1" 304 0 "-" "Python/Requests 22.01"
99.99.44.44 - - [01/01/2021:00:00:25 +0000] "GET /downloads/materialized HTTP/1.1" 304 0 "-" "Python/Requests 22"
91.239.186.133 - - [01/01/2021:00:00:04 +0000] "GET /updates HTTP/1.1" 200 10020 "-" "Python/Requests 22"
173.203.139.108 - - [01/01/2021:00:00:08 +0000] "GET /downloads/materialized HTTP/1.1" 304 0 "-" "Python/Requests 22"
80.91.33.133 - - [01/01/2021:00:00:04 +0000] "GET /downloads/materialized HTTP/1.1" 304 0 "-" "Rust/reqwest 0.3"
93.190.71.150 - - [01/01/2021:00:00:33 +0000] "GET /updates HTTP/1.1" 200 10020 "-" "Python/Requests 22"
91.234.194.89 - - [01/01/2021:00:00:57 +0000] "GET /updates HTTP/1.1" 200 10020 "-" "Python/Requests 22"
46.4.83.163 - - [01/01/2021:00:00:20 +0000] "GET /downloads/materialized HTTP/1.1" 304 0 "-" "Python/Requests 22"
173.203.139.108 - - [01/01/2021:00:00:39 +0000] "GET /downloads/materialized HTTP/1.1" 404 335 "-" "Python/Requests 22"

Example TEXT format

We’ll create a source that ingests all these logs, so that we can execute some quick and dirty analysis:

CREATE MATERIALIZED SOURCE frontend_logs
FROM S3 DISCOVER OBJECTS MATCHING 'logs/202?/**/*.log' USING BUCKET SCAN 'frontend'
WITH (region = 'us-east-2')
FORMAT TEXT;

With that, we will have one SQL row per line in both files, so we can for example query for all the lines that match updates, ordered by position that we encountered the line:

> SELECT mz_record, text FROM frontend_logs WHERE text LIKE '%updates%' ORDER BY mz_record;
 mz_record |                                     text
 1         | 99.99.44.44 - - [12/31/2020:23:55:59] "GET /updates HTTP/1.1" 200 10020 "-" "Python/Requests_22"
 7         | 37.26.93.214 - - [12/31/2020:23:55:52] "GET /updates HTTP/1.1" 200 3318 "-" "Go_1.1_package_http"
 8         | 99.99.44.44 - - [01/01/2021:00:00:41] "GET /updates HTTP/1.1" 200 10020 "-" "Python/Requests_22"
 9         | 188.138.60.101 - - [01/01/2021:00:00:48] "GET /updates HTTP/1.1" 200 10020 "-" "Python/Requests_22"
 12        | 91.239.186.133 - - [01/01/2021:00:00:04] "GET /updates HTTP/1.1" 200 10020 "-" "Python/Requests_22"
 15        | 93.190.71.150 - - [01/01/2021:00:00:33] "GET /updates HTTP/1.1" 200 10020 "-" "Python/Requests_22"
 16        | 91.234.194.89 - - [01/01/2021:00:00:57] "GET /updates HTTP/1.1" 200 10020 "-" "Python/Requests_22"

Example REGEX format

It’s generally more convenient to work with well-structured columnar data, though, so we can use the REGEX format specifier to extract useful data:

CREATE MATERIALIZED SOURCE frontend_logs
FROM S3 DISCOVER OBJECTS MATCHING 'logs/202?/**/*.log' USING BUCKET SCAN 'frontend'
WITH (region = 'us-east-2')
FORMAT REGEX '(?P<ip>[^ ]+) - - \[?P<dt>([^]]_)\] "(?P<method>\w+) (?P<path>[^ ]+)[^"]+" (?P<status>\d+) (?P<content_length>\d+) "-" "(?P<user_agent>[^"]+)"';

With that, we will have one SQL row per line in both files, so we can for example query for all the lines that have /updates as their exact path:

> SELECT dt, ip, user_agent FROM frontend_logs WHERE path = '/updates';
       dt           |      ip        |    user_agent
01/01/2021:00:00:04 | 91.239.186.133 | Python/Requests 22
01/01/2021:00:00:33 | 93.190.71.150  | Python/Requests 22
01/01/2021:00:00:41 | 99.99.44.44    | Python/Requests 22
01/01/2021:00:00:48 | 188.138.60.101 | Python/Requests 22
01/01/2021:00:00:57 | 91.234.194.89  | Python/Requests 22
12/31/2020:23:55:52 | 37.26.93.214   | Go 1.1 package http
12/31/2020:23:55:59 | 99.99.44.44    | Python/Requests 22

Example Single Object

When creating a source that ingests a single object, we only require the ListObject permission:

CREATE MATERIALIZED SOURCE single_object
FROM S3 DISCOVER OBJECTS MATCHING 'logs/2020/12/31/frontend.log' USING BUCKET SCAN 'frontend'
WITH (region = 'us-east-2')
FORMAT TEXT;

Related Pages

• S3 with CSV/JSON encoded data
• CREATE SOURCE
• CREATE VIEW
• SELECT