Introduction

Taba is a service for aggregating instrumentation events from large distributed systems in near real-time. It was built to handle high throughput and scale easily.

Check out an overview of Taba's architecture on the TellApart Eng Blog: http://www.tellapart.com/taba-low-latency-event-aggregation

Example

Taba helps you instrument your services and provide a near real-time view into what's happening across a large cluster. For example, you could use it to track the winning bid price for a certain type of bid:

from taba import client

...

client.Counter('bids_won', 1)
client.Counter('winning_bid_price', wincpm)

When those Events reach the Taba Server and are aggregated, they produce an output like the following:

$ taba-cli agg winning_bid_price
winning_bid_price: {
    "1m": {"count": 436, "total": 571.64},
    "10m": {"count": 5285, "total": 6884.57},
    "1h": {"count": 34265, "total": 44175.47},
    "1d": {"count": 569787, "total": 744423.87},
    "pct": [0.09, 0.47, 2.19, 3.55, 4.37, 14.09, 17.59]}

There are many other input data types and aggregation methods available. See the Types and Handlers documentation.

Overview

A Taba deployment consists of 6 layers, each horizontally scalable. These layers are:

• Taba Client code integrated into applications
• Taba Agent process running locally to the application servers
• Taba Server processes in the frontend ('fe') role
• Taba Server processes in the backend ('be') role
• Redis sentinel processes
• Redis database processes

The Taba Client is integrated into the application it is instrumenting, and exposes an API for recording events to different counter types. The Python distribution includes a default Client implementation based on threads, and a Gevent engine. There is also a Java Client available ([https://github.com/tellapart/taba-java-client] (https://github.com/tellapart/taba-java-client))

The Client typically sends events to a Taba Agent process running on the same server. While the Client will only forward events on a best-effort basis, the Agent provides more robust buffering and failure recovery. It is also significantly smarter about load balancing.

The will forward events to one of the Taba Server end-points it has been configured to connect to. Any Server in the cluster can receive any set of events.

The Taba Server processes are split into two groups: frontend ('fe') and backend ('be'). These roles are configured when the process starts. Assigning a process a 'fe' role has no effect on its operation -- it is simply a marker that the process will use to advertise its role. (The intention it to allow a load-balancer to use that indicator to route traffic to just the 'fe' processes). Assigning the 'be' role will configure the process to launch a background worker that processes queued events. There must be at least one 'be' Server process in the cluster.

A Server process can be assigned both 'fe' and 'be' roles. For small clusters, this will work well. However, once a cluster becomes large enough to require multiple Server processes, separating 'fe' and 'be' processes will perform better.

There is a third role 'ctrl', which essentially marks a Server process as neither 'fe' nor 'be'. This is useful for maintaining a separate set of processes for querying.

The Taba Server uses a group of Redis databases and Sentinels. Having at least one Sentinel is a requirement, as it is used for service discovery of the individual database processes. Sharding across the databases is accomplished by splitting the key-space into a large number of virtual buckets, and assigning ranges of buckets to each process.

Installing Taba

Taba was designed to run on Python 2.6/2.7. It has the following Python package dependencies, which should be installed automatically:

• gevent (>= 0.13.1)
• python-cjson (>= 1.0.5)
• cython (>= 0.13)
• redis (>= 2.9)
• requests (>= 1.2.0)

Additionally, building the Python dependencies requires the following. These dependencies are not installed automatically:

• gcc
• make
• python-dev
• libevent-dev

The latest stable release can be installed from PyPi:

pip install taba

Or Taba can be installed directly from the repository:

git clone https://github.com/tellapart/taba.git
cd taba
python setup.py install

Installing Redis

The Taba Server uses a group of Redis instances with Sentinels as its database. It requires at least Redis 2.8. For details about installing Redis, please visit the Redis Downloads page

Deploying Taba

There are many ways to deploy Taba, depending on the use case. See examples/EXAMPLES for pointers on how to get started.

About

Taba is a project at TellApart led by Kevin Ballard to create a reliable, high performance platform for real-time monitoring. It is used to monitor over 30,000 Tabs, consuming nearly 10,000,000 Events per second, and an average latency of under 15s.

Any questions or comments can be forwarded to taba@tellapart.com