The most enjoyable part of my job at AppDynamics is to witness and evangelize customer success. What’s slightly strange is that for this to happen, an application has to slow down or crash.

It’s a bittersweet feeling when End Users, Operations, Developers and many Businesses suffer application performance pain. Outages cost the business money, but sometimes they cost people their jobs–which is truly unfortunate. However, when people solve performance issues, they become overnight heroes with a great sense of achievement, pride, and obviously relief.

To explain the complexity of managing application performance, imagine your application is 100 haystacks that represent tiers, and somewhere a needle is hurting your end user experience. It’s your job to find the needle as quickly as possible! The problem is, each haystack has over half a million pieces of hay, and they each represent lines of code in your application. It’s therefore no surprise that organizations can take days or weeks to find the root cause of performance issues in large, complex, distributed production environments.

End User Experience Monitoring, Application Mapping and Transaction profiling will help you identify unhappy users, slow business transactions, and problematic haystacks (tiers) in your application, but they won’t find needles. To do this, you’ll need x-ray visibility inside haystacks to see which pieces of hay (lines of code) are holding the needle (root cause) that is hurting your end users. This X-Ray visibility is known as “Deep Diagnostics” in application monitoring terms, and it represents the difference between isolating performance issues and resolving them.

For example, AppDynamics has great End User Monitoring, Business Transaction Monitoring, Application Flow Maps and very cool analytics all integrated into a single product. They all look and sound great (honestly they do), but they only identify and isolate performance issues to an application tier. This is largely what Business Transaction Management (BTM) and Network Performance Management (NPM) solutions do today. They’ll tell you what and where a business transaction slows down, but they won’t tell you the root cause so you can resolve the issues.

Why Deep Diagnostics for Production Monitoring Matters

A key reason why AppDynamics has become very successful in just a few years is because our Deep Diagnostics, behavioral learning, and analytics technology is 18 months ahead of the nearest vendor. A bold claim? Perhaps, but it’s backed up by bold customer case studies such as Edmunds.com and Karavel, who compared us against some of the top vendors in the application performance management (APM) market in 2011. Yes, End User Monitoring, Application Mapping and Transaction Profiling are important–but these capabilities will only help you isolate performance pain, not resolve it.

AppDynamics has the ability to instantly show the complete code execution and timing of slow user requests or business transactions for any Java or .NET application, in production, with incredibly small overhead and no configuration. We basically give customers a metal detector and X-Ray vision to help them find needles in haystacks. Locating the exact line of code responsible for a performance issue means Operations and Developers solve business pain faster, and this is a key reason why AppDynamics technology is disrupting the market.

Below is a small collection of needles that customers found using AppDynamics in production. The simple fact is that complete code visibility allows customers to troubleshoot in minutes as opposed to days and weeks. Monitoring with blind spots and configuring instrumentation are a thing of the past with AppDynamics.

Needle #1 – Slow SQL Statement

Industry: Education
Pain: Key Business Transaction with 5 sec response times
Root Cause: Slow JDBC query with full-table scan

Needle #2 – Slice of Death in Cassandra

Industry: SaaS Provider
Pain: Key Business Transaction with 2.5 sec response times
Root Cause: Slow Thrift query in Cassandra

Needle #3 – Slow & Chatty Web Service Calls

Industry: Media
Pain: Several Business Transactions with 2.5 min response times
Root Cause: Excessive Web Service Invocation (5+ per trx)

Needle #4 -Extreme XML processing

Industry: Retail/E-Commerce
Pain: Key Business Transaction with 17 sec response times
Root Cause: XML serialization over the wire.

Needle #5 – Mail Server Connectivity

Industry: Retail/E-Commerce
Pain: Key Business Transaction with 20 sec response times
Root Cause: Slow Mail Server Connectivity

 Needle #6 – Slow ResultSet Iteration

Industry: Retail/E-Commerce
Pain: Several Business Transactions with 30+ sec response times
Root Cause: Querying too much data

Needle #7 – Slow Security 3rd Party Framework

Industry: Education
Pain: All Business Transactions with > 3 sec response times
Root Cause: Slow 3rd party code

Needle #8 – Excessive SQL Queries

Industry: Education
Pain: Key Business Transactions with 2 min response times
Root Cause: Thousands of SQL queries per transaction

Needle #9 – Commit Happy

Industry: Retail/E-Commerce
Pain: Several Business Transactions with 25+ sec response times
Root Cause: Unnecessary use of commits and transaction management.

Needle #10 – Locking under Concurrency

Industry: Retail/E-Commerce
Pain: Several Business Transactions with 5+ sec response times
Root Cause: Non-Thread safe cache forces locking for read/write consistency

 Needle #11 – Slow 3rd Party Search Service

Industry: SaaS Provider
Pain: Key Business Transaction with 2+ min response times
Root Cause: Slow 3rd Party code

 Needle #12 – Connection Pool Exhaustion

Industry: Financial Services
Pain: Several Business Transactions with 7+ sec response times
Root Cause: DB Connection Pool Exhaustion caused by excessive connection pool invocation & queries

Needle #13 – Excessive Cache Usage

Industry: Retail/E-Commerce
Pain: Several Business Transactions with 50+ sec response times
Root Cause: Cache Sizing & Configuration

If you want to manage and troubleshoot application performance in production, you should seriously consider AppDynamics. We’re the fastest growing on-premise and SaaS based APM vendor in the market right now. You can download our free product AppDynamics Lite or take a free 30-day trial of AppDynamics Pro – our commercial product.

Now go find those needles that are hurting your end users!

App Man.

apm, appdynamics, AppDynamics Pro, application monitoring, Application Performance Management, BTM, Business Transaction Management, CA Wily, Compuware, Deep Diagnostics, Dynatrace, End User Monitoring, New Relic, OpNet, OpTier, Production Monitoring, Root Cause Analysis, Transaction Profiling

Imagine you’re an operations guy and you’ve just received a phone call or alert notifying you that the application your responsible for is running slow. You bring up your console, check all related processes, and notice one java.exe process isn’t using any CPU but the other Java processes are.  The average sys admin at this point would just kill and restart the Java process, cross their fingers, and hope everything returns back to normal (this actually does work most of the time). An experienced sys admin might perform a kill -3 on the Java process, capture a thread dump, and pass this back to dev for analysis. Now suppose your application returns back to normal–end users stop complaining, you pat yourself on the back and beat your chest, and basically resume what you were doing before you were rudely interrupted.

The story I’ve just told may seem contrived, but I’ve witnessed it several times with customers over the years. The stark reality is that no one in operations has the time or visibility to figure out the real business impact behind issues like this. Therefore, little pressure is applied to development to investigate data like thread dumps so that root causes can be found and production slowdowns can be avoided again in future. It’s true restarting a JVM or CLR will solve a fair few issues in production, but it’s only a temporary fix over the real problems that exist within the application logic and configuration.

Now imagine for one minute that operations could actually figure out the business impact of production issues, along with identifying the root cause, and communicate this information to Dev so problems could be fixed rapidly. Sounds too good to be true, right? Well, a few weeks ago an AppDynamics customer did just that and the story they told was quite compelling.

Code Deadlock in a distributed E-Commerce Application

The customer application in question was a busy e-commerce retail website in the US. The architecture was heavily distributed with several hundred application tiers that included JVMs, LDAP servers, CMS server, message queues, databases and 3rd party web services. Here is a quick glimpse of what that architecture looked like from a high level:

Detecting Code Deadlock

If we look at the AppDynamics problem pane (right) as the customer saw things, it shows the severity of their issues. During the day the application was experiencing just over 4,000 business transactions per minute, which works out at just under 1 million transactions a day. Approximately 2.5% of these transactions were impacted by the slowdown, which was the result of the 92 code deadlocks you see here that occurred during peak hours.

AppDynamics is able to dynamically baseline the performance of every business transaction type before classifying each execution as normal, slow, very slow or stalled depending on its deviation from its unique performance baseline. This is critical for understanding the true business impact of every issue or slowdown because operations can immediately see how many user requests were impacted relative to the total requests being processed by the application.

From this pane, operations were able to drill down into the 92 code deadlocks and see the events that took place as each code deadlock occurred. As you can see from the screenshot (below left), the sys admins during the slowdown kept restarting the JVMs (as shown) to try and make the issues go away. Unfortunately, this didn’t work given that the application was experiencing high concurrency under peak load.

By drilling into each Code Deadlock event, operations were able to analyze the various thread contentions and locate the root cause of the issue. The root cause of the slowdown turned out to be an application cache which wasn’t thread-safe. If you look at the screenshot below, showing the final execution of the threads in deadlock accessing the cache, you can see that one thread was trying to remove an item, another was trying to get an item, and the last thread was trying to put an item. 3 threads were trying to do a put, get and remove at the same time! This caused a deadlock to occur on cache access, thus causing the related JVM to hang until those threads were released via a restart.

 Analyzing Thread Dumps

Below you can see the thread dump that AppDynamics collected for one of the code deadlocks, which clearly shows where each thread was deadlocked. By copying the full thread dumps to clipboard, operations were able to see the full stack trace of each thread, thus identifying which business transactions, classes, and methods were responsible for cache access.

The root cause for this production slowdown may have been identified and passed to dev for resolution, but the most compelling conclusion from this customer story was related to them identifying the real business impact that occurred. The application was clearly running slow, but what did the end user experience during the slowdown and what impact would this have had on the business?

What was the Actual Business Impact?

The screenshot below shows all business transactions that were executing on the e-commerce web application during the five hour window before, during, and after the slowdown occurred.

Here are some hard hitting facts for the two most important business transactions inside this e-commerce application:

• 46,463 Checkouts processed
• 482 returned an error, 1325 were slow, 576 were very slow and 111 stalled.
• 3,956 Payments processed
• 12 returned an error, 242 were slow, 96 were very slow and 79 stalled

Error – transaction failed with an exception. Slow – the business transaction deviated from its baseline by more than 3 standard deviations. Very Slow – the business transaction deviated from its baseline by more than 4 standard deviations. Stalled – the transaction timed out.

If you take these raw facts and assume the average revenue per order is $100, then the potential revenue risk/impact of this slowdown was easily into six digits when you consider the end user experience for checkout and payment. Even if you take the 482 Errors and 111 Stalls relating to the Checkout business transaction alone – this still equates to around $60,000 of revenue at risk. And that’s a fairly conservative estimate!

If you add up all the errors, slow, very slow and stalls you see in the screenshot above, you start to picture how serious this issue was in production. The harsh reality is that incidents like this happen everyday in production environments, but no one has visibility into the true business impact of them, meaning little pressure is applied to development to fix “glitches.”

Agile isn’t about Change, It’s about Results

If development teams want to be truly agile, they need to forget about constant change and focus on what impact their releases has on the business. The next time your application slows down or crashes in production, ask yourself one question: “What impact did that just have on the business?” I guarantee just thinking about that answer will make you feel cold. If development teams found out more often the real business impact of their work, they’d learn pretty quickly how fast, reliable and robust their application code really is.

I’m pleased to say no developers were injured or fired during the making of this real-life customer story; they were simply educated on what impact their non-thread safe cache had on the business. Failure is OK–that’s how we learn and build better applications.

App Man.

apm, appdynamics, Application Performance Management, BTM, Business Transaction Management, Code Deadlock, Java Monitoring, Resolving deadlock, Thread Dumps

Ask any developer to name a few common performance bottlenecks and 99.9% of them will say “the database” as their first answer. Ask a DBA the same question and 99.9% of them will respond “crappy application code.” If we put aside the mutual love between each party, it’s actually worth looking at why they arrive at different conclusions.

First of all, it’s very easy to make any application or SQL query perform fast in a dev environment, especially on a developer’s desktop. Why? The two things that impact query performance are data volume and user/transaction concurrency, both of which typically don’t exist in dev environments when unit tests pass. Computing power these days is also silly fast and cheap with desktop cpu and memory helping any dev environment mask potential performance bottlenecks.

Read the Full Post…

App Man, Performance Bottleneck, Slow Database, slow SQL, Slow SQL Statement