Permalink
Switch branches/tags
Nothing to show
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
486 lines (343 sloc) 21.1 KB

Gitter

trck: Query engine for TrailDB

trck is a tool to query TrailDBs for aggregate metrics based on individual user behavior.

Typical use cases:

  • Calculating behavior-based KPIs (like bounce rates)
  • Attributions
  • Extracting features from discrete time series data for machine learning

Example queries:

  • find number of cases where there is no event of type page view within 5 seconds of an event of type click
  • find all user sessions that "bounced" for campaign X
  • find number of user sessions that spend more than 5 minutes on a website

Table of contents

Overview

trck is a domain specific language that defines a finite state machine1 to find patterns in data. These programs are compiled into highly optimized parallel native code.

trck also includes a number of higher level, data-aware optimizations to make processing as efficient as possible, e.g. by looking ahead at the data and skipping parts that have no chance matching a condition altogether.

Data Model

The data type trck programs deal with is a sequence of timestamped events: a log file or a browsing history is a common example. trck uses TrailDB as a storage engine that is optimized for storing large collections of such histories, in a way that makes it easy to process them one by one.

Of course, one can always use TrailDB API directly to analyze that data; however that is often cumbersome and error prone when you need to match complex patterns in trails. Especially when you need to keep track of current matching state for lots of trails simultaneously, when your data is split across multiple TrailDBs.

State Machines

trck approach is to explicitly define a state machine which is then executed for each trail separately. There is no global state you need to keep track of, you only focus on one trail at a time.

A trck program can be considered a device that receives events one by one and reacts to them by transitioning to a different state.

Compiler pipeline

The output of trck program is generated by actions attached to state transitions. For example, you can increment a counter every time state machine transitions from state A to state B.

Optimizations

State machines are compiled to efficient machine code to ensure maximum performance. On top of that, trck compiler applies a few high-level optimizations to skip processing parts of the trails when it has no effect on computation result. With other optimizations, like compressing states to maximize cache efficiency and multicore support, it makes possible for trck programs to process millions of trails per second.

Testing

Matching trail patterns reliably can be very tricky because of a large number of edge cases; that's why trck has a built-in unit test framework and a quickcheck-style property based testing library.

Installation

Requirements:

  • Ubuntu / Debian Linux / OSX
  • Debian packages bash python (>=2.7) make python-ply jq libjudy-dev (>=1.0.5-5) libjson-c-dev libcmph-dev libc6-dev libjemalloc-dev
  • C compiler, tested with gcc and clang.

On OSX, you'll need to install a few packages from brew:

brew install traildb/judy/judy msgpack jemalloc llvm

You'll also need to install TrailDB. As of time of writing, it is recommended to install TrailDB from source as the brew version has a multithreading bug.

Building

git submodule update --init --remote --recursive
make

How to run

Use trck command line compiler from bin/.

Example:

./bin/trck -c myprogram.tr -o matcher-traildb

That will compile your program to a binary matcher-traildb that will dynamically linked to libtraildb. That binary accepts TrailDB paths as positional arguments and prints results in JSON format to stdout. You can also compile a static binary by using --static flag (currently Linux only).

You can specify program parameter values using --params file.json, JSON file should contain a dictionary specifying values for every parameter. See Parameters section for more details.

You can specify output format using --output-format json|msgpack. Currently only single result mode is supported for msgpack output; that means that you have to use merged results mode if you use foreach loops (see below).

Filters

You can also apply filters to traildb to select events a cookies to process. There are two kinds of filters:

  • field filters. You can set field filter by passing --filter flag to a compiled trck program. Filter format TBD. Same filter will be applied to every trail.
  • time window filters. You can pass a path to a csv file using --window-file flag for a compiled trck program. Every line of the file contains 3 comma separated items: uuid, start_timestamp and end_timestamp. For every trail with specified uuid, events having timestamp that doesn't satisfy start_timestamp <= X <= end_timestamp are ignored. Trails that don't have an entry in the file are ignored entirely.
  • uuid exclude filters. You can pass a path to a plain file using --exclude-file for a compiled trck program. Every line of the file must contain a uuid. UUIDs found on this file will be ignored.

Multicore support

trck programs are naturally highly parallelizable. Programs are compiled with OpenMP automatically, if available.

You can explicitly enable or disable OpenMP support using --use-openmp and --no-openmp flags to trck compiler. You can control the number of OpenMP threads at run time for a compiled trck program by setting OMP_NUM_THREADS environment variable, as usual with OpenMP programs, and by default it will use one thread per core.

trck language syntax

Matching rule specification

Matching rules describe state machines in a text form. State machines naturally consist of states and rules describing transitions between states.

Normally state transitions are triggered by events in the trail; therefore transition rules are boolean conditions that match event fields.

In trck you can also trigger transitions on timeouts; they trigger when machine stays in state X for N seconds, without consuming any events.

Basic matching

The syntax is similar to Erlang/Prolog but with indentation-based block syntax aka off-side rule. If you are familiar with Erlang, matcher program is essentially a series of tail recursive message receiver loops, where "messages" are events in the trail.

trck program consists of one or more blocks. Block has a name and a body, separated by keyword receive. Body consists of multiple pattern matching clauses that are evaluated sequentially. Incoming event must match one of the clauses, and the corresponding action is executed.

[block_name] ->
    receive
        ... [condition] -> [action] ...
        ... [condition] -> [action] ...
        ...
    after [timeout] -> [action]

Example:

main ->
    receive
        type = "click", campaign_id = "A"-> yield $clicks
        * -> repeat

This program calculates the number of "click" events for a campaign. receive block can also be seen as an infinite loop, for every arriving event we have two pattern matching clauses here. Action here is the yield statement, that increments a counter variable $clicks every time event in a trail matches the condition.

repeat action just proceeds to the next event, and matching process starts again from the first pattern matching clause.

Clauses must be exhaustive: if none of them match, you get a runtime error. Normally you'd have catch-all *->repeat clause as the last one.

Clause syntax

Let's look closer at the matching clause syntax.

...
    type = "click", campaign_id = "A"-> yield $clicks, skip_dup
...

The part before the -> arrow specifies matching conditions. A comma means boolean AND. There is no OR operator as of now, but you can emulate it by having multiple clauses.

Part after -> arrow is clause action. It may contain yield statements to produce results (like counting events). It also has to contain information on where to go next: it can be either repeat to continue with the next event in this block, or it can be a block name like skip_dup above. In that case, matching will continue with the next event in the trail in the specified block. Another special action is quit that terminates state machine, in case if you don't want to process any more events in this trail.

Block duration and after example

You can also limit receive block duration by specifying windows using after clause. In this example, we'll again count "click" events but will also skip 3 seconds after every click event we encounter, to discard potential duplicates.

main ->
    receive
        type = "click", campaign_id = "A"-> yield $clicks, skip_dup
        * -> repeat
skip_dup ->
    receive
        * -> repeat
    after 3s -> main

Here, after the first clause in main matches, we yield a result, and then go to skip_dup block that just loops over until after 3 seconds, then goes back to the main block.

after clauses cannot use repeat as a transition, and naturally there is no matching condition besides time window specification. Time window can be specified as number and time unit (day/hour/minute/second), e.g. 31h, 3d, 16m, 53s.

Nested windows

You can nest blocks within a parent window block. This is useful when you need to use multiple blocks to implement a matching pattern, but there is "global" expiration period that needs to cover multiple blocks.

The syntax is window ... after:

parentblock1 ->
    window
        block1 ->
            ... clauses ...
        block2 ->
            ... clauses ...
    after 30d -> someblock

Nested blocks can contain transitions to blocks that are outside the window, but not vice versa.

I.e. this is legal, block1 can transition to foo:

...
foo ->
  ...

parentblock1 ->
    window
        block0 ->
            ... clauses ...
        block1 ->
            type = "X" -> foo
        block2 ->
            ... clauses ...
    after 30d -> someblock

But you can't go from foo to block1, this code won't compile:

foo ->
    receive
        type = "X" -> block1
        * -> repeat

parentblock1 ->
    window
        block0 ->
            ... clauses ...
        block1 ->
            ... clauses ...
        block2 ->
            ... clauses ...
    after 30d -> someblock

Parameters

Jobs may use parameters that are passed externally to them in matching conditions:

main ->
  receive
      type = "click", campaign_id = %id -> yield $clicks
      type = "click", campaign_id in #foo -> yield $clicks
      * -> repeat

%id and #foo can be passed at run time to the compiled program. Parameters can be scalars like above, or sets of values. Set parameter names start with #, scalar names start with %.

You can only compare parameters to field values, not other parameters or constants. That limitation allows the compiler to apply very aggressive optimization techniques. Note that parameters also have (implicit) field "types", e.g. you cannot compare %id to campaign_id field value in one clause and compare it to user_id field value in another. You can work around this by simply having two separate parameters.

Multiple parameter group matching with foreach

For batch processing, you often need to run the program for a number of values; for example, finding number of clicks for every campaign. That, of course, could be done by making campaign id a parameter and running program N times for every parameter value. However, that may be pretty inefficient especially if you have more than a few values.

That's where foreach statement is useful. You can create a top level for-loop that loops through values and executes the program for every value, then separate result set is produced for each iteration. That is, in above example, every campaign would get a separate conversion counter. This is the same as if you ran your program N times, once for each campaign, except performance is much better.

Syntax is

foreach %bar,#something,%foo in @param
   block1 ->
       receive
           ....
   ...

Here @param is specified in parameter config file (see --params option for compiled binary), and contains a list of tuples, each item type should match the corresponding variable.

In the above example, first and third items of every tuple must be strings and the second item is a set (encoded in JSON as array):

{
  "@param" : [
    ["A", ["item1", "item2"], "B"]
    ["Z", ["item1", "item3"], "C"]
  ]
}

If you use quit action within foreach statement, it will only stop trail processing for this specific parameter set (tuple).

One can think of foreach as a state machine cloning operation; without foreach you have a separate instance of your state machine for each trail, with separate state. With foreach and N tuples as parameters, you'll instead have N instances of your state machine for every trail. Therefore quit statement only terminates one of these state machines.

Implicit foreach arrays

Sometimes you just want to run foreach over a simple list of all values of some field occurring in the traildb. For example, run your query over all possible campaigns. Since extracting all of them manually and saving to a parameters config file can be a bit tedious -- all required information is already in traildbs anyway -- there is a shortcut for this case:

foreach %cid
   block1 ->
       receive
           campaign_id = %cid, ... -> ...
           ....
   ...

You can omit array parameter name in the foreach clause, in that case, matcher will extract list of possible field values from traildbs and will loop over them.

Returning results: yield statements

yield is used to return results from the program. As discussed in the previous section, a separate result is returned for each iteration of foreach. Meaning that if you want to run your program over multiple configurations passed in foreach array, you'll get a separate result for each configuration.

Merged results

If you want trck to automatically merge these results, you can add merged results qualifier to your foreach loop:

foreach %aeid,%seid in @arr merged results
 ...

Counting things: yield to a counter

Simplest form of yield is

yield $counter

Where $counter is an output variable. This statement would increase counter named $counter by one. There is no need to declare output variables in advance.

Example:
foreach %cid in @arr
    start ->
        receive
            type = "cli", campaign_id = %cid -> yield $clicks
            * -> repeat

This example would produce a separate counter for each item in @arr, e.g.:

[
  {"%cid" : "A", "$clicks" : 5},
  {"%cid" : "B", "$clicks" : 3}
]

Counting things: yield field to a set

Another form of yield allows you to return field values as a tuple. Syntax is

yield field1,field2 to #result

Where field1, field2 are field names and #result is an output variable. #result is a set: that is, yielding the same tuple of fields multiple times will produce only one copy in #result. In JSON format as returned by trck, variable #result would be a list of strings, each string containing field set separated by commas.

Counting things pt II: yield field to a multiset

Very similar to sets, can be used in the case when you don't only want distinct items but also their counts. Instead of producing an array of distinct items, you get a dictionary that contains items as keys and counts as values.

yield field1,field2 to &result

In JSON format as returned by trck, variable &result would look like this:

...
"&result" : { "foo" : 1, "bar" : 2}
...

Counting things: yield field to HLL (cardinality estimator)

Information on HLLs can be found here Sets and multisets can also be yielded as an HLL data structure encoded in hex using the syntax:

yield field1,field2 to ^result

In JSON format as returned by trck, variable ^result would look like this:

...
"^result": "0e01a7040001028738000101f80f0001049a21000101bc1200"
...

The HLL is encoded as follows:

 0e 01 a7040001028738000101f80f0001049a21000101bc1200
 -- -- ----------------------------------------------
 ^  ^                       ^
 |  |                       |
(1)(2)                     (3)

1) HLL precision (p), bins = 2^p
2) HLL version:
    0 = Empty
    1 = Non-empty
3) HLL run-length encoded bins, RLE pairs can be stored in two or three bytes,
   if the MSB is set, then two bytes were used to store the length.

Complete example:

foreach %aeid in @arr
    start ->
        receive
            type = "cli", advertisable_eid = %aeid -> yield segment_eid,ad_eid to #ads
            type = "imp", advertisable_eid = %aeid -> yield domain to &domains
            type = "pxl", advertisable_eid = %aeid -> yield cookie to ^cookies
            * -> repeat

As described above, this example produces a separate result for each item in @arr, e.g.:

[
    {"%aeid" : "A", "#ads" : ["seg1,ad1", "seg2,ad2"], "&domains" : {"example.com" : 12, "news.org" : 41}, "^cookies": "0e01a7040001028738000101f80f0001049a21000101bc1200"},
    {"%aeid" : "B", "#ads" : ["seg1,ad2"], "&domains" : {"example.com" : 84, "news.org" : 11}, "^cookies": "0e01d128000102ba06000102b63e000101bc1200"}
]

Calling user defined functions in yield

You can call external functions when yielding to a set or a multiset; just drop a file with the same name as your .tr script but with .tr.c extension and you'll be able to call functions from that file in yield statements. See test_ffi.tr in test/tr directory for an example.

Complete example:

foreach %cid in @arr
    start ->
        receive
            type = "cli", campaign_id = %cid -> yield segment_eid,ad_eid to #ads
            type = "imp", campaign_id = %cid -> yield domain to &domains
            * -> repeat

As described above, this example produces a separate result for each item in @arr, e.g.:

[
  {"%cid" : "A", "#ads" : ["seg1,ad1", "seg2,ad2"], "&domains" : {"example.com" : 12, "news.org" : 41} },
  {"%cid" : "B", "#ads" : ["seg1,ad2"], "&domains" : {"example.com" : 84, "news.org" : 11}}
]

If you used merged results mode, you'll get a single result dictionary:

foreach %cid in @arr merged results
    start ->
        receive
            type = "cli", campaign_id = %cid -> yield segment_eid,ad_eid to #ads
            type = "imp", campaign_id = %cid -> yield domain to &domains
            * -> repeat

Produces following output instead:

{"#ads": ["seg1,ad1", "seg2,ad2", "seg1,ad2"],
 "&domains": {"example.com": 96, "news.org": 52}}

Special variables

In addition to returning values of traildb fields in yield ... to statements, you can also return special variables:

  • timestamp contains current event timestamp in the trail
  • cookie contains current cookie as a hexadecimal string
  • start_timestamp contains timestamp when current receive block was entered. Right now it is only possible to use this variable if that block has an after clause, i.e. block cannot be infinite.
  • start_timestamp[label] contains timestamp when parent window block with the name label was entered. It is only possible to use this variable within the referred block.

Implementation details

Compiler pipeline

A program first parsed and converted to intermediate JSON-based format. trparser.py does this and it is fairly independent of the rest of the pipeline.

Second, fsm2c.py produces a C module and a header from that JSON representation. That contains state machine itself. That part is technically independent of the underlying storage and can be used with a storage engine other than traildb or even online event processing system that doesn't store entire trails.

Then C module produced by fsm2c.py is compiled with match_traildb.c which provides necessary primitive functions that are called by generated code to access data from traildbs and also implements efficient foreach block execution, setting parameters, collecting results and printing them.

1 arguably not a finite state machine in a strict sense, but close enough

Contributors