Skip to content

Latest commit

 

History

History
533 lines (439 loc) · 19.5 KB

README.md

File metadata and controls

533 lines (439 loc) · 19.5 KB

angle-grinder Build Status Gitter chat

Slice and dice log files on the command line.

Angle-grinder allows you to parse, aggregate, sum, average, min/max, percentile, and sort your data. You can see it, live-updating, in your terminal. Angle grinder is designed for when, for whatever reason, you don't have your data in graphite/honeycomb/kibana/sumologic/splunk/etc. but still want to be able to do sophisticated analytics.

Angle grinder can process well above 1M rows per second (simple pipelines as high as 5M), so it's usable for fairly meaty aggregation. The results will live update in your terminal as data is processed. Angle grinder is a bare bones functional programming language coupled with a pretty terminal UI.

overview gif

Quick Links

Installation

Binaries are available for Linux and OS X. Many more platforms (including Windows) are available if you compile from source. In all of the commands below, the resulting binary will be called agrind. Starting with v0.9.0, agrind can self-update via the --self-update flag.

With Brew (OS X)

brew install angle-grinder

With Curl (Single binary)

Linux:

curl -L https://github.com/rcoh/angle-grinder/releases/download/v0.12.0/angle_grinder-v0.12.0-x86_64-unknown-linux-musl.tar.gz \
  | tar Ozxf - \
  | sudo tee /usr/local/bin/agrind > /dev/null && sudo chmod +x /usr/local/bin/agrind

OS X:

curl -L https://github.com/rcoh/angle-grinder/releases/download/v0.12.0/angle_grinder-v0.12.0-x86_64-apple-darwin.tar.gz \
  | tar Ozxf - \
  | sudo tee /usr/local/bin/agrind > /dev/null && sudo chmod +x /usr/local/bin/agrind

From Source

If you have Cargo installed, you can compile & install from source: (Works with Stable Rust >=1.26)

cargo install ag

Query Syntax

An angle grinder query is composed of filters followed by a series of operators. The filters select the lines from the input stream to be transformed by the operators. Typically, the initial operators will transform the data in some way by parsing fields or JSON from the log line. The subsequent operators can then aggregate or group the data via operators like sum, average, percentile, etc.

agrind '<filter1> [... <filterN>] | operator1 | operator2 | operator3 | ...'

A simple query that operates on JSON logs and counts the number of logs per level could be:

agrind '* | json | count by log_level'

Filters

There are three basic filters:

  • *: Match all logs
  • filter-me* (with no quotes) is a case-insensitive match that can include wildcards
  • "filter-me" (in quotes) is a case-sensitive match (no wildcards, * matches literal * , filter-me, or "filter me!".

Filters can be combined with AND, OR and NOT

("ERROR" OR WARN*) AND NOT staging | count

Sub-expressions must be grouped in parenthesis. Only lines that match all filters will be passed to the subsequent operators. filter.gif

Aliases

Starting with v0.12.0, angle grinder supports aliases, pre-built pipelines do simplify common tasks or formats. The only alias currently defined is apache, which parses apache logs. Adding more aliases is one of the easiest ways to contribute!

Examples:

* | apache | count by status

Operators

Non Aggregate Operators

These operators have a 1 to 1 correspondence between input data and output data. 1 row in, 0 or 1 rows out.

JSON

json [from other_field]: Extract json-serialized rows into fields for later use. If the row is not valid JSON, then it is dropped. Optionally, from other_field can be specified. Nested JSON structures are supported out of the box. Simply access nested values with .key[index], for example, .servers[6]. Negative indexing is also supported.

Examples:

* | json
* | parse "INFO *" as js | json from js

Given input like:

{"key": "blah", "nested_key": {"this": "that"}}
* | json | count_distinct(nested_key.this)

json.gif

Logfmt

logfmt [from other_field]: Extract logfmt-serialized rows into fields for later use. If the row is not valid logfmt, then it is dropped. Optionally, from other_field can be specified. Logfmt is a an output format commonly used by Heroku and Splunk, described at https://www.brandur.org/logfmt.

Examples:

* | logfmt

Given input like:

{"key": "blah", "nested_key": "some=logfmt data=more"}
* | json | logfmt from nested_key | fields some
Split

split[(input_field)] [on separator] [as new_field]: Split the input via the separator (default is ,). Output is an array type. If no input_field or new_field, the contents will be put in the key _split.

Examples:

* | split on " "

Given input like:

INFO web-001 influxd[188053]: 127.0.0.1 "POST /write HTTP/1.0" 204

Output:

[_split=[INFO, web-001, influxd[188053]:, 127.0.0.1, POST /write HTTP/1.0, 204]]

If input_field is used, and there is no new_field specified, then the input_field will be overridden with the split data-structure. For example:

* | parse "* *" as level, csv | split(csv)

Given input like:

INFO darren,hello,50
WARN jonathon,good-bye,100

Will output:

[csv=[darren, hello, 50]]        [level=INFO]
[csv=[jonathon, good-bye, 100]]        [level=WARN]

Other examples:

* | logfmt | split(raw) on "blah" as tokens | sum(tokens[1])
Parse

parse "* pattern * otherpattern *" [from field] as a,b,c [nodrop]: Parse text that matches the pattern into variables. Lines that don't match the pattern will be dropped unless nodrop is specified. * is equivalent to regular expression .* and is greedy. By default, parse operates on the raw text of the message. With from field_name, parse will instead process input from a specific column. Any whitespace in the parse expression will match any whitespace character in the input text (eg. a literal tab).

Examples:

* | parse "[status_code=*]" as status_code

parse.gif

Fields

fields [only|except|-|+] a, b: Drop fields a, b or include only a, b depending on specified mode.

Examples: Drop all fields except event and timestamp

* | json | fields + event, timestamp

Drop only the event field

* | fields except event
Where

where <bool-expr>: Drop rows where the condition is not met. The condition must be an expression that returns a boolean value. The expression can be as simple as a field name or a comparison (i.e. ==, !=, <=, >=, <, >) between fields and literal values (i.e. numbers, strings). The '!' operator can be used to negate the result of a sub-expression. Note that None == None, so a row where both the left and right sides match a non-existent key will match.

Examples

* | json | where status_code >= 400
* | json | where user_id_a == user_id_b
* | json | where url != "/hostname"
Limit

limit #: Limit the number of rows to the given amount. If the number is positive, only the first N rows are returned. If the number is negative, the last N rows are returned.

Examples

* | limit 10
* | limit -10
Field Expression

<expr> as <name>: The given expression is evaluated and the result is stored in a field with the given name for the current row. The expression can be made up of the following:

  • +, -, *, /: Mathematical operators with the normal precedence rules. The operators work on numeric values and strings that can automatically be converted to numbers.
  • ==, !=, <=, >=, <, >: Boolean operators work on most data types.
  • <field>: The name of a field in the current row. If the row does not contain the given field, an error will be reported.
  • Parentheses to group operations

The following functions are supported within expressions:

  • Mathematical functions: abs(), acos(), asin(), atan(), atan2(), cbrt(), ceil(), cos(), cosh(), exp(), expm1(), floor(), hypot(), log(), log10(), log1p(), round(), sin(), sinh(), sqrt(), tan(), tanh(), toDegrees(), toRadians()
  • concat(arg0, ..., argN) - Concatenate the arguments into a string
  • contains(haystack, needle) - Return true if the haystack contains the needle.
  • length(str) - Returns the number of characters in "str".
  • parseDate(str) - Attempt to parse a date from the given string.
  • substring(str, startOffset, [endOffset]) - Returns the part of the string specified by the given starting offset up to the end offset (if specified).

Examples Multiply value by 100 to get the percentage

* | json | value * 100 as percentage

Aggregate Operators

Aggregate operators group and combine your data by 0 or more key fields. The same query can include multiple aggregates. The general syntax is:

(operator [as renamed_column])+ [by key_col1, key_col2]

In the simplest form, key fields refer to columns, but they can also be generalized expressions (see examples) Examples:

* | count
* | json | count by status_code
* | json | count, p50(response_ms), p90(response_ms) by status_code
* | json | count as num_requests, p50(response_ms), p90(response_ms) by status_code
* | json | count, p50(response_ms), p90(response_ms), count by status_code >= 400, url

There are several aggregate operators available.

Count

count[(condition)] [as count_column]: Counts the number of input rows. Output column defaults to _count. Optionally, you can provide a condition -- this will count all rows for which the condition evaluates to true.

Examples:

Count number of rows by source_host:

* | count by source_host

Count number of source_hosts:

* | count by source_host | count

Count the number of info vs. error logs:

* | json | count(level == "info") as info_logs, count(level == "error") as error_logs
Sum

sum(column) [as sum_column]: Sum values in column. If the value in column is non-numeric, the row will be ignored. Examples:

* | json | sum(num_records) by action
Min

min(column) [as min_column] [by a, b] : Compute the min of values in column. If the value in column is non-numeric, the row will be ignored.

Examples:

* | json | min(response_time)
Average

average(column) [as average_column] [by a, b] : Average values in column. If the value in column is non-numeric, the row will be ignored.

Examples:

* | json | average(response_time)
Max

max(column) [as max_column] [by a, b] : Compute the max of values in column. If the value in column is non-numeric, the row will be ignored.

Examples:

* | json | max(response_time)
Percentile

pXX(column): calculate the XXth percentile of column

Examples:

* | json | p50(response_time), p90(response_time) by endpoint_url, status_code
Sort

sort by a, [b, c] [asc|desc]: Sort aggregate data by a collection of columns. Defaults to ascending.

Examples:

* | json | count by endpoint_url, status_code | sort by endpoint_url desc

In addition to columns, sort can also sort an arbitrary expressions.

* | json | sort by num_requests / num_responses
* | json | sort by length(endpoint_url)
Total

total(a) [as renamed_total]: Compute the running total of a given field. Total does not currently support grouping!

Examples:

* | json | total(num_requests) as tot_requests
Count Distinct

count_distinct(a): Count distinct values of column a. Warning: this is not fixed memory. Be careful about processing too many groups.

Examples:

* | json | count_distinct(ip_address)

Example Queries

  • Count the number of downloads of angle-grinder by release (with special guest jq)
curl  https://api.github.com/repos/rcoh/angle-grinder/releases  | \
   jq '.[] | .assets | .[]' -c | \
   agrind '* | json
         | parse "download/*/" from browser_download_url as version
         | sum(download_count) by version | sort by version desc'

Output:

version       _sum
-----------------------
v0.6.2        0
v0.6.1        4
v0.6.0        5
v0.5.1        0
v0.5.0        4
v0.4.0        0
v0.3.3        0
v0.3.2        2
v0.3.1        9
v0.3.0        7
v0.2.1        0
v0.2.0        1
  • Take the 50th percentile of response time by host:
tail -F my_json_logs | agrind '* | json | pct50(response_time) by url'
  • Count the number of status codes by url:
tail -F  my_json_logs | agrind '* | json | count status_code by url'

More example queries can be found in the tests folder

Rendering

Non-aggregate data is simply written row-by-row to the terminal as it is received:

tail -f live_pcap | agrind '* | parse "* > *:" as src, dest | parse "length *" as length'
[dest=111.221.29.254.https]        [length=0]        [src=21:50:18.458331 IP 10.0.2.243.47152]
[dest=111.221.29.254.https]        [length=310]      [src=21:50:18.458527 IP 10.0.2.243.47152]

Alternate rendering formats can be provided with the --output flag. Options:

  • --output json: JSON output
  • --output logfmt: logfmt style output (k=v)
  • --output format=<rust formatter>: This flag uses rust string formatting syntax. For example:
    tail -f live_pcap | agrind --format '{src} => {dst} | length={length}' '* | parse "* > *:" as src, dest | parse "length *" as length'
    21:50:18.458331 IP 10.0.2.243.47152 => 111.221.29.254.https | length=0
    21:50:18.458527 IP 10.0.2.243.47152 => 111.221.29.254.https | length=310
    

Aggregate data is written to the terminal and will live-update until the stream ends:

k2                  avg
--------------------------------
test longer test    500.50
test test           375.38
alternate input     4.00
hello               3.00
hello thanks        2.00

The renderer will do its best to keep the data nicely formatted as it changes and the number of output rows is limited to the length of your terminal. Currently, it has a refresh rate of about 20hz.

The renderer can detect whether or not the output is a tty -- if you write to a file, it will print once when the pipeline completes.

Contributing

angle-grinder builds with Rust >= 1.26. rustfmt is required when submitting PRs (rustup component add rustfmt).

There are a number of ways you can contribute:

  • Defining new aliases for common log formats or actions
  • Adding new special purpose operators
  • Improve documentation of existing operators + providing more usage examples
  • Provide more test cases of real queries on real world data
  • Tell more people about angle grinder!
cargo build
cargo test
cargo install --path .
agrind --help
... write some code!

cargo fmt

git commit ... etc.

When submitting PRs, please run cargo fmt -- this is necessary for the CI suite to pass. You can install cargo fmt with: rustup component add rustfmt if it's not already in your toolchain.

See the following projects and open issues for specific potential improvements/bugs.

Project: Improving Error Reporting

Usability can be greatly improved by accurate and helpful error messages for query-related issues. If you have struggled to figure out why a query is not working correctly and had a hard time fixing the issue, that would be a good place to jump in and start making changes!

First, you need to determine where the problem is occurring. If the parser is rejecting a query, the grammar may need some tweaking to be more accepting of some syntax. For example, if the field names are not provided for the parse operator, the query can still be parsed to produce a syntax tree and the error can be raised in the next phase. If the query passes the parsing phase, the problem may lie in the semantic analysis phase where the values in parse tree are verified for correctness. Continuing with the parse example, if the number of captures in the pattern string does not match the number of field names, the error would be raised here. Finally, if the query has been valid up to this point, you might want to raise an error at execution time. For example, if a field name being accessed does not exist in the records being passed to an operator, an error could be raised to tell the user that they might have mistyped the name.

Once you have an idea of where the problem might lie, you can start to dig into the code. The grammar is written using nom and is contained in the lang.rs module. The enums/structs that make up the parse tree are also in the lang.rs module. To make error reporting easier, values in the parse tree are wrapped with a Positioned object that records where the value came from in the query string. The Positioned objects are produced by the with_pos!() parser combinator. These objects can then be passed to the SnippetBuilder in the errors.rs module to highlight portions of the query string in error messages.

The semantic phase is contained in the typecheck.rs module and is probably where most of the work will need to be done. The semantic_analysis() methods in that module are passed an ErrorBuilder that can be used to build and send error reports to the user.

After adjusting the grammar and adding a check for the problem, it will be time to figure out how to inform the user. Ideally, any errors should explain the problem, point the user to the relevant part of the query string, and lead the user to a solution. Using the ErrorBuilder, you can call the new_error_report_for() method to construct a SnippetBuilder for a given error. To highlight a portion of the query string, use the with_code_pointer() method with the Positioned object that refers to the relevant segment of the query string. Finally, additional help/examples can be added by calling the with_resolution() method.

Once you're all done, you should see a nicely formatted error message like the following:

error: Expecting an expression to count
  |
1 | * | count_distinct
  |     ^^^^^^^^^^^^^^ No field argument given
  |
  = help: example: count_distinct(field_to_count)

Similar Projects

  • Angle Grinder is a rewrite of Sumoshell written to be easier to use, testable and a better platform for new features.
  • lnav is a full featured log analysis platform in your terminal (with many more features than angle-grinder). It includes support for common log file formats out-of-the-box, generalized SQL queries on your logs, auto-coloring and a whole host of other features.
  • visidata is a spreadsheets app in your terminal