Secret Techniques from the Toughest Tableau Server Deployments

Tamas Foldi 

Starschema, CTO

tfoldi@starschema.com - @tfoldi

Who am I?

OK, so large deployments?

How do you survive with this complexity?

Automate.
Analyse.
Extend.

Automation

Basic AWS Tableau Infrastructure

Tons of manual processes

Problem

Human work does not scale well.

Error prone

Upgrades, test and dev systems deploymens are not agile enough

Lack of integration with modern CI/CD pipelines 

Solution

Automate all processes with infrastructure provision scripts

Do not perform in-place upgrades: always build new from scratch and drop the old. Seriously. 

Automate!

Automate using

• CloudFormation scripts
• Chef, Ansible, Terraform
• Docker
• ANYTHING JUST
  DON'T DO MANUALLY

Docker Demo

Communication

Communication vs Tableau

Problem

Tableau does not have a good in-built communication system where administrators and CoE leaders can access their users

No out-of-the-box solution to access users in case of system outages

Solution

Add custom communication capabilities to Desktop and Server products

Custom Firewall and Loadbalancer scripts

Outage Notification
(Demo)

Notification Best Practices

• Custom Load Balancer Setup
• "Trust" or "Status" site
• Eliminate Emails
  (use IM or in-vizportal)

CoE Communication

Customized Desktop

Understanding the infrastructure

Understand? What? Why?

Problem

Administrators usually have zero control on what gets published or where their servers connects. Getting answers to "what are my badly designed workbooks?", "what happens if a team changes their data model?" or "Is my Tableau Server fast enough for my overpaid manager?" can help to survive.

TabMon and many other tools do not scale.

Solution

Watch serverlogs, infrastructure events, repository and workbook/datasource XML sources and build sophisticated preventive and detective controls on top of them. 

HOW?

The new serverlogs

New "res" block 

What is inside?

alloc: memory 

free: free'd up memory

ucpu: CPU consumption (application side)

kcpu: CPU consumption (kernel side)

ntid: used threads

e = exclusive (self), i = inclusive (self + children),  ne = number of exclusive calls, ni = number of inclusive calls.

Real-time, proactive alerting

But what about the metadata?

Lineage, database tables and columns are in the XML files

Metadata directly available in the Dashboard

DEMO

How to understand your users?

Tableau Tracker Demo

Advanced Topics

Off-site backup

Problem

Traditional Tableau backups can quickly grow into hundreds of GB-s and take hours to create on your primary node.

Backup frequency will necessary go down and if a problem occurs between backups multiple day’s worth of data might be lost.  

Solution

Off-load backup to external computers: stream postgres repository outside the cluster and replicate filestore. Build tsbak files on separate machine. This also supports point-in-time recovery.

Disaster Recovery /
Multi-region Replicatoin

Traffic Management

Problem

CEO spends 2$m on data visualisation then cries that her reports are slow

Backgound jobs fail to make their SLA

Web Authoring slows down the system when users are experimenting  

Solution

Isolate unpredictable load (QA site, interactive users, special set of background jobs) to dedicated hosts/processes

Route specific user groups to dedicated resources 

Traffic Management

Questions?

Tamas Foldi

tfoldi@starschema.com / twitter: @tfoldi