TrailDB is a portable C library that allows querying a series of relative events. It is used to group the existing relative events in a time-series format and produce an immutable database with high compression rate.
It is designed as a complement to current existing relational databases or key-value stores and targeted for OLAP workload such as analyzing usage patterns, predicting user behavior, and detecting anomalies. One key design feature is that the database is immutable once produced. This immutability feature enables data compression. TrailDB leverages relativity among time-series events to achieve high compression. These two key features allow TrailDB to achieve good performance in OLAP workload.
TrailDB system has been developed at AdRoll to handle the user-level analytics. The developers at AdRoll has seen the increasing number of requests for user-level analytics but found it hard to grasp and analyze a large user-level data using advanced SQL queries. Thus, they developed TrailDB system that is used to query a series of events at user-level.
The first version was released May 2016. The next version released in May 2017 introduced indexes and a new query engine (trck).
TrailDB system does not support checkpoints as each database is immutable once produced.
First, within a trail, events are always sorted by time. Thus, it utilizes Delta Encoding to compress the 64-bit timestamps.
Second, since events are grouped by
UUID, which usually represents a logical entity such as an online shopping customer, these events within a trail tend to be predictable and TrailDB only encodes every change in behavior. This is not exactly the same as the Run-Length Encoding but similar.
Third, Huffman Coding, which is a kind of Prefix Compression method, is used to encode the skewed, low-entropy distributions of values.
As each TrailDB is an immutable file, modifications are not allowed. There is only one process to produce a database and no one can issue read operations before the creation is finalized. Thus, concurrency is not needed in TrailDB system.
TrailDB system adopts a specific relational data model. The traditional relational data model consists of a key and a set of different attributes. In TrailDB system, it consists of a key called
UUID and a list of objects consists of values of a set of pre-defined fields.
TrailDB system defines a data entity called a trail that is uniquely identified by a
UUID. Within each trail, there is a list of ordered events, each of which is identified and ordered by the timestamp. For each event, it contains values for the pre-defined set of fields. These fields are similar to attributes in the traditional relational data model.
This data model allows the relative events belonging to one
UUID, taking one online shopping user as an example, to group together in the order of time. Thus, it offers the predictability feature among the list of events and enables TrailDB system developers to use several compression methods to achieve high compression rate and extraction speed in TrailDB.
In TrailDB, each database consists of a collection of trails each of which is identified by a unique
UUID. There are no multiple tables within a database and no constraints among databases. Thus, it does not support foreign keys.
TrailDB indexes use a specific inverted index to map each item to a list of page ids. Each item is uniquely identified by a field and the value in that field.
Indexes map items to a list of page ids that contains that item. There is a file contains a
FIELD SECTION. TrailDB system looks into the
HEADER first to get the field's corresponding beginning offset of
FIELD SECTION. Then, it finds out the corresponding item and extracts the page ids containing that item.
When creating a database, there's only one process to handle it and others cannot access it. Once the database is produced, it is a read-only immutable file. Thus, everyone can issue read requests to it, but cannot issue any write operations. In this point of view, it is equivalent to the serializable isolation level.
TrailDB system offers the APIs to allow a join operation on multiple trails. Within each trail, events are already sorted in timestamp order. TrailDB system leverages this feature and adopts the merge sort of multiple trails to produce one single merged trail with a list of sorted events.
TrailDB does not support logging and there's only one process to create the database. There is no recovery handler if the process crashes during the creation of the database. Thus, users need to start from the very beginning of the producing process.
But, TrailDB system allows merging existing TrailDBs to create a new immutable database. It is suggested to do so if there's a huge number of input events.
TrailDB system offers custom APIs to allow users to query events with cursors. It can emit one event each time with one cursor or multiple events with multiple cursors. There are next functions provided to move the cursor(s) to the next event(s) in the trail(s).
TrailDB system offers its own custom APIs instead of using the standard SQL query interface. TrailDB is designed to make a specific organization of the user-level events and it is born not fit for SQL. Instead, It offers the query interface in several programming languages: C, Go, Python, R, Haskell, and D. TrailDB system also provides two specific languages to query the data. One is called trck, which is a domain specific language to aggregate metrics based on events of identical
UUID. The other is called Reel, a small general query language for TrailDB.
Each TrailDB is a read-only immutable file, it does not support stored procedures.
The TrailDB adopts the embedded database architecture. Each database created by TrailDB system is an immutable file. Thus, everyone can have a copy of this database and access it using the custom API or specific query languages as a standalone application. There's no administrator needed for this database.
TrailDB system does not support views. But, as each database is an immutable file, users can create "views" by creating another immutable database by extracting data from the existing TrailDBs.