2 releases

0.1.1 Jun 6, 2022
0.1.0 Feb 27, 2020

#1880 in Database interfaces


Used in liquid-ml

MIT license

180KB
898 lines

SoRer

SoRer, short for schema-on-read-er, is a library that can infer a schema, parse .sor files into a columnar representation according to the schema, and handle missing data and (most cases of) malformed data.

SoRer was built with speed and memory efficiency in mind and file parsing is multi-threaded.

On our 2 year old desktop computer with a SATA SSD (meaning our testing is likely near being bottlenecked by ssd read speeds) and 4 cores (4 threads), SoRer can parse at ~400 MB/s on a large test file with 8 columns, two of each data type with random values (which can be generated by running cargo run --release --bin generate (warning don't do this inside of Docker, you must install rust if you want to do this due to file i/o overhead when using Docker). In a best case scenario, on a large file with 3 columns of random bools, it can parse at over 700 MB/s

Usage

Building SoRer

SoRer can be built on any computer by running the command: make docker from the root of this repository. This builds a Docker image tagged as sorer. It also builds the executable for sorer, located at /sorer/target/release/sorer and copies over the executable to the current directory.

Tests can be ran by running the command make test. The program can be ran against a small test file named sor.txt by running the command: make run.

Documentation can be built by running the command make doc. This builds the documentation and copies it to ./doc/ on the host filesystem in this directory. This documentation can be viewed by opening ./doc/sorer/index.html in your broswer.

Note that ideally the best way to run our program is bare metal due to overhead for using Docker (especially on Windows or Mac). You can do that by installing rust by running the following command:

curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh

Follow the printed instructions to source the cargo environment variables after installing.

Then build sorer by running cargo build --release. You may test the program after installing rust by running cargo test. Documentation may be built by running cargo doc --no-deps --open.

Running SoRer

SoRer is ran as a command line tool that prints its results to stdout.

The command line arguments are summarized in the below table

argument value type required? description
-f <string> yes path to SoR file
-from <uint> no starting position in file (in bytes)
-len <uint> no number of bytes to read
-print_col_type <uint> depends print the type of a column: BOOL, INT, FLOAT, STRING
-print_col_idx <uint> <uint> depends the first argument is the column, the second is the offset
-is_missing_idx <uint> <uint> depends is there a missing field in the specified column offset

When <val> in -from <val> is greater than 0, then the file is read starting from the first complete line after <val>.

When <val> in -len <val> is greater than 0, then the file is read up until the last complete line.

After running make build, running make bash will mount the current the current directory to the docker container and start bash. If you want to test any large files, you should do make build first, then copy the files into this directory, then run make bash. Once you're in bash, you can interact with sorer as usual:

SoR Files

A SoR file is stored as plain text. Files consists of a sequence of rows, each row must be separated by the newline character, "\n". Each row is a sequence of fields, each field starting with "<" and ending with ">". Spaces around delimiters are ignored.

SoR Fields

A field can be either missing a value, or contain a value of one of four SoR types:

  • String
  • Float
  • Integer
  • Bool
Type Allowed values
String Either as a sequences of characters without spaces or as a double quote delimited sequence of characters with spaces. Line breaks are not allowed in Strings. Can't be longer than 255 characters. Must be valid utf-8 characters.
Float Any C++ float
Integer Any C++ integer, ie a sequence of digits with an optional leading sign (must not be separated by whitespace)
bool {1, 0}
Missing (aka Null) must be empty, ie "<>"

Valid Examples of SoR Fields

The following is an example of a row with four fields:

< 1 > < hi >< +2.2 > < " bye ">

The following is an example of a row with explicit missing fields:

<1> <bye> <> <>

The following is also valid:

<> <> <> <>

Invalid Examples of SoR Fields

<1. 2>       // space after dot

<bye world>  // string with spaces and without quotes

<+ 1>        // space after the +

NOTE: If a SoR file contains an invalid field, the row will be discarded for both schema inference and data parsing.

Schema Inference

The schema that SoRer generates depends on the data types contained in the row with the most number of fields in the first 100 rows, followed by 100 rows from the mid-point of the file, and finally with the final 100 rows (or the whole file, whichever comes first). In the sorer example, these rows are used irregardless of the --from command line argument. The data type chosen for each column in the schema is the highest-precedence data type that was seen in all the rows that were equal to the width of the widest row.

The Data Type precedence is as follows:

  1. String
  2. Float
  3. Integer
  4. Bool

This means that if any value is a String, the whole column is parsed into a String type. Otherwise, if any of the values is a Float, then the column is of Float type. Otherwise, if you find a value with a sign or a value larger than 1, then the column is Integer. Otherwise the column is a Bool type, even if there were only explicit 'missings' and no data.

Rows that don't match the schema

If a row that doesn't match the schema is found after the schema is inferred (meaning after the first 500 lines), then the row is discarded. An example is if a schema is parsed as <int> <int>, but a line coming after the first 500 has <string> <int>, then it will be discarded.

Note however, that it is valid for two rows in the same file to have a different number of fields and still be considered to match the schema. For rows with more fields than the schema, the extra fields will be discarded but the row will still be parsed as long as the other fields match the schema.

E.g. The schema: <int> <bool> and a row: <12> <0> <discarded> parses to <12><0>

If a row has less fields without explicit missing fields (i.e. "<>"), aka implicit missing fields, SoRer will attempt to parse the fields according to the schema and fill in explicit missing fields at the end of the row until it matches the number of fields in the schema.

E.g. The schema: <int> <bool> <string> and a row: <12> parses to <12><><>

pub mod dataframe; pub mod parsers; pub mod schema;

Dependencies

~2.3–3.5MB
~61K SLoC