What it’s like to use Haskell

By Andy Friesen

Since early 2013, we at IMVU have used Haskell to build several of the REST APIs that power our service.

When the company started, we chose PHP as our application server language, in part, because the founders expected the website to only be a small part of the business!  IMVU was primarily about a downloadable 3D client.  We needed “a website or something” to give users a place to download our client from, but didn’t expect it would have to be much more than that. This shows that predicting the future is hard.
Years later, we have quite a lot of customers, and we primarily use PHP to serve them.  We’re big enough that we run multiple subteams on separate initiatives at the same time.  Performance is becoming important to us not just because it matters to our customers, but because it can easily make the difference between buying 4 servers and buying 40 servers to support some new feature.

So, early in 2012, we found ourselves ready to look for an alternative that would help us be more rigorous.  In particular, we were ready for the idea that sacrificing a tiny bit of short term, straight-line time to market might actually speed us up in the long run.

How We Got Here

I started learning Haskell in my spare time in part because Haskell seems like the exact opposite of PHP: Natively compiled, statically typed, and very principled.

My initial exploration left me interested in evaluating Haskell at real scale.  A year later, we did a live-fire test in which we taught multiple teammates Haskell while delivering an important new feature under a deadline.

Today, a lot of our backend code is still driven by PHP, but we have a growing amount of Haskell that powers newer features. The process has been exciting not only because we got to actually answer a lot of the questions that keep many people from choosing not to try Haskell, but also because it’s simply a better solution.

The experiment to start developing in Haskell took a lot of internal courage and dedication, and we had to overcome a number of, quite rational, concerns related to adopting a whole new language. Here are the main ones and how they worked out for us:

Scalability

The first thing we did was to replace a single service with a Haskell implementation.  We picked a service that was high-volume but was not mission critical.

We didn’t do any particular optimization of this new service, but it nevertheless showed excellent performance characteristics in the field.  Our little Haskell server was running on a pair of spare servers that were otherwise set for retirement, and despite this, each machine was handling about 20x as many requests as one of our high-spec PHP servers could manage.

Reliability

The second thing we did was to take our hands off the Haskell service and leave it running until it fell over.  It ran for months without intervention.

Training

After the reliability test, we were ready to try a live fire exercise, but we had to wait a bit for the right project.  We got our chance in early 2013.

The rules of the experiment were simple: Train 3 engineers to write the backend for an important new project and keep up with a separate frontend team.  Most of the code was to be new, so there was relatively little room for legacy complications.

We very quickly learned that we had also signed up for a lot of catch-up work to bring the Haskell infrastructure inline with what we’ve had for years in PHP.  We were very busy for awhile, but once we got this infrastructure out of the way, the tables turned and the front-end team became the limiting factor.

Today, training an engineer to be productive in our Haskell code is not much harder than training someone to be productive in our PHP environment.  People who have prior functional programming knowledge seem to find their stride in just a few days.

Testing

Correctness is becoming very important for us because we sometimes have to change code that predates every current developer.  We have enough users that mistakes become very costly, very quickly.  Solving these sorts of issues in PHP is sometimes achievable but always difficult.  We usually solve them with unit tests and production alerts, but these approaches aren’t sufficient for all cases.

Unit tests are incredible and great, but you’re always at the mercy of the level of discipline of every engineer at every moment. It’s easy to tell your teammates to write tests for everything, but this basically boils down to asking everyone to be at their very best every day.  People make mistakes and things slip through the cracks.

When using Haskell, we actually remove an entire class of defects that we have to write tests for. Thus, the number of tests we have to write is smaller, and thus there are fewer cases we can forget to write tests for.

We like unit testing and test-driven development (TDD) at IMVU and we’ve found that Haskell is better with TDD, but also that TDD is better with Haskell.  It takes fewer tests to get the same degree of reliability out of Haskell.  The static verification takes care of quite a lot of error checking that has to be manually implemented (or forgotten) in PHP.  The Haskell QuickCheck tool is also a wonderful help for developers.
The way Haskell separates pure computations from side effects let us build something that isn’t practical with other languages: We built a custom monad that lets us “switch off” side effects in our tests.  This is incredible because it means that trying to escape the testing sandbox breaks compilation. While we have had to fight intermittent test failures for eight years in PHP (and at times have had multiple engineers simultaneously dedicated to the problem of test intermittency,) our unit tests in Haskell cannot intermittently fail.

Deployment

Deployment is great. At IMVU, we do continuous deployment, and Haskell is no exception. We build our application as a statically linked executable, and rsync it out to our servers. We can also keep old versions around, so we can switch back, should a deployment result in unexpected errors.

I wouldn’t write an OS kernel in it, but Haskell is way better than PHP as a systems language. We needed a Memcached client for our Haskell code, and rather than try to talk to a C implementation, we just wrote one in Haskell.  It took about a half day to write and performs really well. And, as a side effect, if we ever read back some data we don’t expect from memcached (say, because of an unexpected version change) then Haskell will automatically detect and reject this data.

We’ve consistently found that we unmake whole classes of bugs by defining new data types for concepts to wrap primitive types like integers and strings.  For instance, we have two lines of code that say that “customer IDs” and “product IDs” are represented to the hardware as numbers, but they are not mutually convertible.  Setting up these new types doesn’t take very much work and it makes the type checker a LOT more helpful. PHP, and other popular dynamic server languages like Javascript or Ruby, make doing the same very hard.

Refactoring is a breeze.  We just write the change we want and follow the compile errors.  If it builds, it almost certainly also passes tests.

Not All Sunshine and Rainbows

Resource leaks in Haskell are nasty.  We once had a bug where an unevaluated dictionary was the source of a space leak that would eventually take our servers down.  We also ran into an issue where an upstream library opened /dev/urandom for randomness, but never closed the file handle.  These issues don’t happen in PHP, with its process-per-request model, and they were more difficult to track down and resolve than they would have been in C++.

The Haskell package manager, Cabal, ended up getting in the way of our development. It lets you specify version ranges of particular packages you want, but it’s important for everyone on the team to have exactly the same versions of every package.  That means controlling transitive dependencies, and Cabal doesn’t really offer a way to handle this precisely. For a language that is so very principled on type algebra, it’s surprising that the package manager doesn’t follow suit regarding package versioning. Instead, we use Cabal for basic package installation, and a custom build tool (written in Haskell.)

Hiring

I’ll admit that I was very worried that we wouldn’t be able to hire great people if our criteria was expertise in an uncommon language without a comparatively sparse industrial track record, but the honest truth is that we found a great Haskell hacker in the Bay area after about 4 days of looking.

We had a chance to hire him because we were using Haskell, not in spite of it.

Final Thoughts

While it’s usually difficult to objectively measure things like choice of programming language or softwarestack, we’re now seeing fantastic, obvious productivity and efficiency gains.  Even a year later, all the Haskell code we have runs on just a tiny number of servers and, when we have to make changes to the code, we can do so quickly and confidently.

Charming Python: How to actually close a socket by calling shutdown() before calling close()

By Eric Hohenstein

You may never have to use the Python low-level socket interface but if you ever do, here’s a tip regarding some surprising Python socket behavior that might help you.

The standard Python library has a “socket” module written in Python that wraps an underlying “_socket” module written in C that wraps OS level sockets. Together, these expose a cross-platform Berkeley socket-like interface to Python applications. I say Berkeley socket-like because the socket functions are exposed on an object rather than global free-functions. The surprise with this interface is that the close() function of the socket object does not close the socket. Instead, it de-references it, allowing garbage collection to eventually close the OS-level socket. If you only use the close() method on a socket, the socket will continue using both client and server resources until it is garbage collected. Worse, if the other end is continuing to send data to the locally discarded socket, it may block when the receive window of local socket fills up. Even worse than that, if the software handling the other end of the socket is written in erlang and 0 window packets are dropped somewhere in between (for instance because of overly aggressive firewall rules), the other end may block until the socket gets garbage collected even if the socket is configured to immediately time out send operations if they would block.

The solution is to first call shutdown() on the socket before calling close(). You have to be careful doing this if the socket is being read/written by multiple threads simultaneously but down that path lies madness, so don’t do that.

The Python documentation for the socket module actually does briefly mention this but doesn’t really give much of an explanation as to why it’s necessary. Really, the Python documentation on the subject is correct in that calling shutdown is better form. However, in cases where you are sure that the application state of the connection is no longer valid, it should be possible to bypass the shutdown and close the socket to simply free its resources immediately.

Why is it necessary to call shutdown() before close()? Let’s take a look at the socket.py module from the standard Python library. These code snippets are all taken from Python 2.6 source, though I’ve confirmed that Python 2.7 and Python 3.0 behave similarly. First, socket.py imports everything from _socket, which includes a class called socket:

import _socket

from _socket import * 

Then, it saves a reference to _socket.socket called _realsocket

_realsocket = socket

Then it defines a list of functions that will be exposed from the real socket object to the wrapper class (see below):

_socketmethods = (         

          ‘bind’, ‘connect’, ‘connect_ex’, ‘fileno’, ‘listen’,         
          ‘getpeername’, ‘getsockname’, ‘getsockopt’, ‘setsockopt’,         
          ‘sendall’, ‘setblocking’,         
          ‘settimeout’, ‘gettimeout’, ‘shutdown’) 

Note that ‘close’ is not in that list but ‘shutdown’ is.

Then it defines a _closedsocket class that will fail all operaitons:

class _closedsocket(object):   

        __slots__ = []   
        def _dummy(*args):       
                raise error(EBADF, ‘Bad file descriptor’)   
         # All _delegate_methods must also be initialized here.   
          send = recv = recv_into = sendto = recvfrom = recvfrom_into = _dummy   

          __getattr__ = _dummy

Then, it defines a class called _socketobject. Note the comment before the start of this class:

# Wrapper around platform socket objects. This implements

# a platform-independent dup() functionality. The

# implementation currently relies on reference counting

# to close the underlying socket object.

class _socketobject(object):    

       __doc__ = _realsocket.__doc__    
       __slots__ = ["_sock", "__weakref__"] + list(_delegate_methods)    
       def __init__(self, family=AF_INET, type=SOCK_STREAM, proto=0, _sock=None):       
                if _sock is None:           
                      _sock = _realsocket(family, type, proto)       
                self._sock = _sock       
                for method in _delegate_methods:           
                        setattr(self, method, getattr(_sock, method)) 

Within the _socketobject class, it defines a method for each of the functions exposed by the _socket class:

             _s = (“def %s(self, *args): return self._sock.%s(*args)\n\n”         

            “%s.__doc__ = _realsocket.%s.__doc__\n”)   
             for _m in _socketmethods:       
                      exec _s % (_m, _m, _m, _m)   
             del _m, _s 

Remember that close was not in the _socketmethods list. Here’s the close method of the _socketobject class:

             def close(self):       

                     self._sock = _closedsocket()       
                     dummy = self._sock._dummy       
                     for method in _delegate_methods:           
                             setattr(self, method, dummy)   
              close.__doc__ = _realsocket.close.__doc__ 

Assigning a new value to self._sock de-references the real socket object, allowing it to be garbage collected. It then exposes the _socketobject class as socket:

socket = SocketType = _socketobject

Client networking code using the Python socket library would create a socket like so:

import socket

sock = socket.socket() 

which will create a socket._socketobject instance that has a _sock member which is a _socket.socket instance. After connecting the socket to a remote endpoint, calling close() just assigns the _sock member to an instance of _closesocket, allowing Python to eventually garbage collect the original _socket.socket object. After the previous code, the following will “leak” the OS socket for some arbitrary amount of time:

sock.connect((“www.foobar.com“, 80))

sock.close() 

Looking at the C code in _socketmodule.c, there is a close() method defined on the _socket.socket class but there is no code in socket.py that will call it:

static PyObject *

sock_close(PySocketSockObject *s)
{
SOCKET_T fd; 
if ((fd = s->sock_fd) != -1) {
s->sock_fd = -1;
Py_BEGIN_ALLOW_THREADS
(void) SOCKETCLOSE(fd);
Py_END_ALLOW_THREADS
}
Py_INCREF(Py_None);
return Py_None;
}  

Actually, there is code in socket.py that will call this function but it’s in the socket._fileobject class that is returned from the socket._socketobject makefile() function which typical networking code will likely not use.

If the close() method of socket._socketobject deferred to the _socket.socket close() method, calling close() on a socket.socket object would actually close the socket (if it was open). Since it does not, calling shutdown() is the only way to close the socket immediately. I’ll not go into the difference between close() and shutdown() here since it’s fairly complex and mostly unimportant but a simplified explanation is that shutdown() may wait until any unsent data has been flushed to the network (although it may not) and close() will not.

Here’s the implementation of _socket.socket shutdown() function:

static PyObject *

sock_shutdown(PySocketSockObject *s, PyObject *arg)
{
int how;
int res; 
how = PyInt_AsLong(arg);
if (how == -1 && PyErr_Occurred())
return NULL;
Py_BEGIN_ALLOW_THREADS
res = shutdown(s->sock_fd, how);
Py_END_ALLOW_THREADS
if (res < 0)
return s->errorhandler();
Py_INCREF(Py_None);
return Py_None;
}  

Calling shutdown() on an instance of socket._socketobject will actually shutdown the OS level socket before it is garbage collected. If the _socket.socket object is not already shutdown or closed, it will eventually be closed when it’s garbage collected:

 

static void

sock_dealloc(PySocketSockObject *s)
{
if (s->sock_fd != -1)
(void) SOCKETCLOSE(s->sock_fd);
Py_TYPE(s)->tp_free((PyObject *)s);
} 

The conclusion is that Python sockets should always be closed by first calling shutdown() and then calling close().

Efficient and Scalable Off-Site Backup to Amazon Glacier

By Ted Reed

The strength of IMVU is our large catalog of User-Generated Content (UGC). With more than ten million items in our virtual catalog, losing our UGC would be crippling to our business.  We in the Operations Team take the preservation of this data seriously. We recently upgraded our aging UGC backup system to use Amazon Glacier as the storage medium. This post will briefly explain the old backup system before detailing the new one.

In The Beginning

We store our UGC in a MogileFS instance, a system for cheap and efficient storage of files across commodity hardware. Our offsite backups originally took the form of USB drives onto which a process would copy the files as they were written to Mogile. As each drive filled up, we would then transport it from our colocation facility to a fire safe in our office. To cover the period of time when the disk was still being written to, we would keep a synced copy of the disk on a machine in a server closet in our offices. This copy would then be deleted once the USB disk had been safely stored.

As time went on, the rate at which our customers were adding photos or uploading products to our catalog grew and grew. We also started permitting higher-quality photos and made it easier to upload and manage them. In order to compensate for the increase in growth, we started buying larger and larger USB drives. This past summer we reached the point where we had the biggest USB drives we could reasonably get and we were still filling them up in about a week. Each time a drive filled up, one of us had to drive out to the facility to retrieve it.

Auditing the data also became woefully inconvenient, as the number of individual drives one had to juggle during the audit skyrocketed. It was an annoyingly manual process, even with helper scripts to manage everything but plugging and unplugging drives. Additionally, annoyingly manual processes tend to not get done as often as we ought to.

Enter Amazon Glacier

We spent some time looking over options for better off-site backups. We talked to many vendors and even for those who could handle our “monumental amount of data” (actual term used by a vendor who shall remain nameless). The quotes left us wondering if we might not be better off just putting some cheap servers in another data center somewhere.

While we were evaluating these options in late August 2012, Amazon announced Glacier, an “extremely low-cost storage service” intended for long-term archival of infrequently-accessed data. At just one cent per gigabyte per month, it handily beat out every other vendor we’d seen in terms of price. We poked around and tested the service out, and found it to be right up our alley, although the pricing was annoyingly confusing. (I ended up writing a small Haskell program to re-implement the math from their FAQ. It rounds in weird places.)

The Backup Process

The backup process begins with a Perl daemon which manages worker pools of Fetchers and Uploaders. The master process figures out which files need to be backed up and hands work off to the Fetchers, which talk to Mogile and pull the file down to a local directory. Once a directory has about 25 MB of files, it’s closed off and handed to an Uploader. The Uploader will then tar the directory and send the tar to Glacier.

It’s worth pointing out here that the 25 MB bundling is actually pretty important if you don’t want to pay a great deal of money. Glacier charges five cents per thousand uploads. My first tests ignored that cost and we pretty quickly ran up an unexpectedly large bill. After some analysis, we found that 25 MB was about right as a middle point between cost and flexibility. This middle point is likely to differ for your use case and budget. The state of the backup process is stored in a MySQL database, which has tables for archives and the files that go into them. We record roughly the same information that Mogile does about each file so that we can accurately restore it if the file is accidentally deleted from Mogile. For each archive, we store both its size as well as information about where to find it (region/vault/id). We also have tables to record information about backup failures for later investigation.

Restoration

The restoration process begins with a simple front-end script, which can take a source and destination. The source can be in terms of Mogile ID, Mogile Domain/Key, or one of our URLs (the latter two are dereferenced to Mogile ID). The destination can be Mogile itself, a local file, or S3. There’s also a batch mode which takes input where each line represents one source and one destination. The script will record each file to restore in the database. It will then issue requests to Glacier to retrieve the archives needed to restore the files, unless there is already an active and uncompleted request for the same archive.

Elsewhere in our network, we have a daemon which sits and listens to an SQS queue tied to the SNS topic for restoration. The notification will include the database ID for the restore request. The daemon will pull the information needed to do the restoration and then clear it afterwards.

Auditing and Validation

One side benefit of a backup system that doesn’t involve managing disks is that we can automate testing the backups and our restoration process. There are two aspects that we want to test. First, we want to prove that the backups are intact and accurate. Second, we want to prove that we are able to restore the data. Glacier permits you to pull up to 5% of your total data per month, presumably for purposes such as these. You still need to pay bandwidth egress charges if the data leaves AWS, which informed our design for the automated audit process.

In order to audit the backups, there’s a cron on our side which runs hourly and selects a random sampling of one millionth of the archives we’ve uploaded. If we had 3,000,000 archives, we’d audit 3 per hour. For each archive, this cron pulls the files from Mogile and generates md5sums. It issues a request to Glacier, with a different SNS topic than the one we use for ordinary restoration. It then uploads the md5sums as a file to an EC2 micro instance which runs a daemon listening for that SNS topic. As the archives become available, the daemon does the same md5sum generation on the Glacier data, comparing it to the list uploaded beforehand. If there is a discrepancy, it sends an email to a mailing list that we check at least daily.

To handle verifying the restoration path, there is a smaller monthly process. It will generate a list of 10 random files and use our normal restoration tool to restore them to test buckets in both Mogile and S3. If the files haven’t been restored within 24 hours, a notification goes to the same mailing list as for audit failures.

In The End

We’re still pushing our historical data into Glacier, but this project is already looking pretty successful. We’ve been adding new data since last November, and are in the process of checking through everything to gain a full trust in the system before we finally pull the plug on the old system and build a mighty throne out of the USB drives.

The backup process was mostly optimized to be able to push these historical files up to Amazon in a reasonable timeframe. As a result, the process which backs up our real-time uploads is quite fast with plenty of room to grow as our usage scales up. Files added to Mogile are in the backup process within seconds, and are typically in Glacier within less than a minute. I looked at our system during peak time for an example, and found that files were in Glacier 35 seconds after being uploaded to us by a user. Knowing this gives me great peace of mind that our UGC will be safe in the event a disaster strikes.

What about other forms of backup? We briefly looked at storing our offsite database backups in Glacier, but decided against this. Any data you put into Glacier will be billed for at least three months. Adding daily backups and then deleting some of them later makes for a rather expensive solution. There’s also little to gain as the process which moves the database backups to our office isn’t nearly as painful as our older UGC backup process was.

Monitoring Delayed Replication, with a focus on MySQL

By Stuart Cianos, CISSP

Author’s Preface

Many (many) years ago, I had the opportunity to work with a variety of organizations targeting harm reduction for at risk populations in the San Francisco Bay Area. One of the great lessons learned from that experience is that the future is mostly indeterminate. In 1997-1998, severe storms ravaged our region and very few of us dealing with relief services (myself included) anticipated the total impact. It was only by sheer luck and coincidence that there was available shelter space to house an entire community of displaced families after automatic flood gates failed to open on a large stream following years of drought. We didn’t think it would happen to us until it was too late to be better prepared. If I win the lottery tomorrow, my future might change in ways that were unexpected. Conversely, if I were to accidentally delete all of the files (“rm -rf /”) on a mission critical computer system, my future might also change in unexpected ways.

I no longer worry about the future, I prepare for it. What follows is a powerful technique which can help mitigate risks around unanticipated data loss, even when that data loss is due to human error.

Introduction

IMVU utilizes MySQL on a daily basis to drive our site and customer experience forward. As IMVU has grown and scaled, so have the number of database hosts and the importance of data maintained in our environment. With over 100,000,000 registered user accounts and hundreds of thousands of concurrent transactions at any given time, it’s of paramount importance that our practices embody the CIA triad: confidentiality, integrity and accessibility of information. Implementing delayed replication at the data tier can enhance business continuity and improve posture when recovering from events impacting data integrity and availability.

IMVU uses MogileFS to store user generated binary content (images, content creator products, etc.). MogileFS is a user-space fault tolerant, distributed file system. MogileFS stores filesystem metadata in a MySQL database (PostgreSQL is also supported, but not used in our environment). The system is designed to ensure that no single failure will result in data loss or service degradation.

In June of 2012, human error resulted in the MogileFS metadata table (“file”) being dropped from the primary database instance rather than a standby under maintenance. Without any way to relate requests for files back to the content on disk, MogileFS could no longer serve any requests resulting in customers being unable to access some major features of the IMVU social network and rich 3D experience. There was a slave database instance lagging behind its master due to write heavy operations, which allowed us to use it as a recovery source. We were very fortunate that this slave was lagging behind unintentionally; had it not been, the impact of this incident would have been much more severe.

Delayed replication is being implemented across IMVU’s cluster of database instances in order to protect against these types of accidental operations moving forward.

Data Availability and the CIA Triad

The CIA triad is a set of common attributes which should be applied to information management and security. Information on the CIA triad is available abundantly online, but a brief summary is below:

Stuart blog1

  • Confidentiality: Prevention of information disclosure to unauthorized entities.
  • Integrity: Prevention and detection of the unauthorized modification of data.
  • Availability: Assurance that the information will be available when needed.

One question that comes up is what delayed replication has to do with information security… and my response is “a lot!” Delayed replication can help mitigate risks around data availability as well as data integrity, and the impact an incident will have on the bottom line.

What is Database Replication?

Database replication allows transactions on one database host to be replicated to one or more separate instances. There are many replication strategies, but the examples discussed will focus on a simple Master-Slave configuration throughout.

Stuart blog2

When the transaction “INSERT INTO foo VALUES (‘hello world’)” is committed to the master database instance, it is replicated on the slave and committed there as well. By having a slave or standby host available and relatively up to date, it is possible to recover very quickly from various hardware and software failures on the master by replacing it with the slave/standby host.

Stuart blog3

What is Delayed Replication?

Delayed replication is the technique of inserting a delay line into a database’s replication mechanism. Transactions will each be held in a first-in, first-out queue for two hours prior to committal on the slave host being delayed. For the purposes of this example, “mytable” is a mission critical set of data required for the application to function and contains 100 gigabytes of data. If the table “mytable” becomes unavailable for whatever reason (for example, a drop table statement was executed accidentally/unintentionally on the primary and it immediately replicated and got executed on the standby), customers will not have access to the service. In this example, fictitious business parameters will be defined to help us quantify possible benefits:

  • Recovery time objective (RTO): 8 hours
  • Recovery point objective (RPO): 4 hours
  • Single loss expectancy is calculated using the standard formula: SLE=(Asset Value) * (Exposure Factor).
  • Our exposure factor (a subjective percentage of functionality or impact): 100%, since no customers will be able to access the system.
  • Timespan used to calculate the single loss expectancies in terms of hours is the time from failure to service availability.
  • Value of customer transactions against “mytable”, per hour: $10,000 USD
  • Single loss expectancy based on recovery time objective of 8 hours: (10,000 * 8) * (1.00) = $80,000.
  • Total time to rebuild “mytable” from a backup: 20 hours
  • Total time to fail over a database from master to standby/slave: 5 minutes

Given the above, a single loss expectancy can be calculated for “mytable” assuming a worst case scenario of ~20 hours. The realistic single loss expectancy (for the table, not the entire database server asset as a whole) may be calculated: (10,000 * 20) * (1.00) = $200,000. This is $120,000 loss above and beyond what management would have expected. Worse, the recovery time objective and recovery point objectives cannot be consistently met.

Adding a two hour delay line (well within the recovery point objective) can help meet business requirements and reduce single loss expectancy.

stuart blog5

The above transaction seems pretty innocuous, and having a delay line enabled doesn’t make sense for some use cases such as read slaves that are expected to be relatively consistent with their master. So… why bother with a delay line at all?

The benefits become clear when looking at another example. What if the INSERT statement above is changed to DELETE FROM mytable (or even DROP TABLE mytable) due to a bug in code, human error or malicious intent? In an environment with immediate committal on all slaves, one must:

1. Hold their breath and:

◦     Hope that a slave can be stopped before the transaction propagates, dump the table, and restore to the impacted master (likely 8+ hours).
◦     Hope that a slave can be stopped before the transaction propagates and fail over to it.

2. Restore from backup (20 hours).

Hope must not be factored into sound business practices; option one is off the table. With a delay line enabled procedures may be implemented to create a recovery process:

stuart blog4

The recovery process and benefits when using delayed replication can be documented, made repeatable and proven. So long as the problem is caught before the offending transaction makes it through the delay line, the recovery process is:

  1. T+0 minutes: Service degraded. Investigation into cause begins.
  2. T+5 minutes: Cause determined. Stop all replication to the delayed slave (on MySQL, don’t forget to stop the slave IO thread in addition to the SQL thread).
  3. T+35 minutes: Remove the pending transaction in the relay log, or, reset the relay log based on best judgment and the importance of data consistency.
  4. T+40 minutes: Promote the delayed standby to master.
  5. T+50 minutes: Validate that the system is functional.
  6. Service recovery declared.
  7. Recover the demoted master to a consistent state.

The total time to recovery during the above incident was 50 minutes, with an approximate loss of (10,000 * 0.833) * (1.00) = $8,333. The recovery time objective and recovery point objective have been met.

Delayed Replication and Seconds Behind Master:

MySQL 5.6 natively includes delayed replication as a new feature, and will be configurable via the CHANGE MASTER TO statement. MySQL 5.6 is not a generally available release, however, so is not deployed in production for most environments. Organizations running versions 5.1 and 5.5 can effectively implement delayed replication using the Percona-Toolkit utilities (formerly Maatkit) from Percona Software. The Percona Toolkit is an open source collection of helpful tools focused towards management of the MySQL database server. At IMVU, we have successfully implemented delayed replication using the percona-toolkit tools (more specifically, pt-slave-delay).

One side effect of delayed MySQL replication via an external process is that it is no longer possible to fully determine slave state using SHOW SLAVE STATUS. As the slave’s SQL_THREAD is periodically stopped and started, “Seconds behind master” will be NULL most of the time.

In order to mitigate the lack of information from MySQL’s internal replication state, a second utility is available through the percona-toolkit: pt-heartbeat. Pt-heartbeat writes a timestamped entry to a table periodically, creating a heartbeat.

The heartbeat transaction will be replicated, and delayed. By comparing the slave’s current time with the timestamp on the last committed heartbeat, we can approximate with fair certainty the number of seconds the slave has been delayed (or is lagging behind the master).

Monitoring the Solution, effectively:

Monitoring delayed replication becomes more complex than simply watching for heartbeats and comparing timestamps:

  • Usually, the slave will be stopped during backups unless something like Percona Xtrabackup or Enterprise Backup is being used. The replication delay will increase for the duration of the backup, and then contract. It should be noted that IMVU is OK with taking backups from a two hour old data source given our business requirements. Always check with your organization’s business requirements; don’t assume!
  • The replication delay will be affected by MySQL’s replication characteristics, including the fact that it is single threaded. High transaction volume can cause replication delays to increase and subsequently contract. The replication delay is variable and will not be consistent on a heavily loaded database.
  • If the pt-heartbeat/pt-slave-delay fails to maintain the delay line and monitoring, the team must always be made aware.
  • If the replication fails due to IO or SQL thread errors, the team must always be made aware.
  • If the database replication or monitoring system fails for any other reason, the team must be made aware. Contingencies must be in place for failures in MySQL, as well as in code which monitors delayed replication.

Rather than looking at the difference in timestamps at a moment in time, IMVU took the approach of calculating the slope of the delay across a timespan. With the additional information, descriptive statistics are calculated:

  • The slope of the delay line’s values over a time period: Allows determination of whether or not the delay line is stable, moving towards the desired value, or away from it and how quickly.
  • The Y-intercept of the slope: Treated as the approximate current number of seconds behind, functionally equivalent to MySQL’s Seconds Behind Master.
  • Correlation Coefficient: Allows determination of how well the current values on the trend line correlate as a series. If the correlation is 0, there is very little correlation meaning that the values are highly distributed over the given timespan. For values -1 < p < 0, there is correlation between values and the line is trending downward. For values 0 < p < 1, there is correlation between values and the delay line is trending upwards.

Additional information is helpful as well, and can be calculated based on data gathered from MySQL:

  1. The time when backups were kicked off.
  2. The number of seconds backups have been running (if applicable).
  3. The number of seconds the trend line has been in a degraded state due to: backups running, recovering after a backup, and trending towards recovery but not in a backup state.

If backups are being taken from the delayed standby/slave, the delay line will increase dramatically when replication is paused.

To prevent false alarms from paging our operations team, monitoring sensitivity must be decreased during backup windows. Once the backup finishes, it will take time for MySQL’s replication thread to catch up to the desired delay line. Again, sensitivity must be reduced during the recovery window.

If the delay line briefly dips below the configured value for 20 seconds but recovers 30 seconds later, the monitoring system should not page as this would be considered a false alarm at 3:00 in the morning. This is not a real time computing system/database platform, so there is no guarantee that the delay line will maintain at exactly the configured value… only a promise that it will maintain the delay line as close to the target as possible. On average, the delay line varies by up to 10 seconds when not under load and not under impact.

In a mission critical system, any type of exception to standard monitoring should have clearly defined parameters and values around the specific exception trigger, and the limits of the exception. In IMVU’s case, the limits around the exception are based on time. Every exception which results in reduced sensitivity has a corresponding limit to how long the service can remain in that state. If the limit is exceeded, the service goes into an alarm state for Nagios to process. Therefore, no external process may lessen sensitivity of the monitoring beyond configurable limits.

Delayed replication and monitoring in action!

Here is a real example of delayed replication running in a production environment (currently being staged across our cluster). It should be noted that additional exceptions and limits are in place for our internal processing. Our implementation will accept the following configurable limits and ranges:

–max-seconds-behind=<seconds>: Max seconds behind for local heartbeat (there needs to be a frequent heartbeat detected from the local host itself as well as the master – or there’s a problem!)

–max-relay-behind=<seconds>: Max seconds for relay log last update

–max-master-behind=<seconds>: Max seconds for master log last update

–max-backup-behind=<seconds>: Max seconds a backup may run

–max-deshard-behind=<seconds>: Max seconds a deshard job may run

–max-last-seen=<seconds>: Max seconds since worker last seen

–min-delay-time=<seconds>: Minimum allowed replication transaction delay time

–max-delay-time=<seconds>: Maximum allowed replication transaction delay time

–max-recovery-time=<seconds>: Seconds after backup completes to allow variances

–max-trend-time=<seconds>: Maximum time to allow intercept to trend to recovery

The actual limits passed to our Nagios plugin during our staged roll out are currently:

–max-seconds-behind 60

–max-master-behind 1800

–max-relay-behind 28800

–max-backup-behind 64800

–max-deshard-behind 32400

–max-last-seen 60

–min-delay-time 5400

–max-delay-time 9000

–max-recovery-time 21600

–max-trend-time 9000

A database instance in good health with a 2 hour replication delay:

stuart blog6

A database instance currently running an active backup. If the backup was not running, this service would be in an alarm/critical state, as the replication delay line has grown to 16,408 seconds (far above our window limit of 9,000 seconds, and the target goal of 7,200 seconds):

stuart blog7.jpg

Once a backup finishes, the recovery window is entered. If recovery back to the desired delay line does not occur within the recovery window, the service will go critical:

stuart blog8

The complete list of delayed replication conditions trapped at IMVU is listed below, along with the service state (i.e. WARNING, CRITICAL):

  • WARNING: Replication delay intercept recovering – below minimum with slope: Replication delay is below the minimum desired value, but is moving towards recovery.
  • WARNING: Replication delay intercept recovering – above maximum with slope: Replication delay is above the maximum desired value, but is moving towards recovery.
  • WARNING: Replication delay intercept is out of bounds, currently in recovery window for backup which completed <seconds> seconds ago: Replication delay is above/below the desired range, and may not be moving towards recovery. State is held in warning to allow recovery post backup until the recovery window expires.
  • WARNING: Replication delay intercept is out of bounds, currently in recovery window for deshard which completed <seconds> seconds ago: Replication delay is above/below the desired range, and may not be moving towards recovery. State is held in warning to allow recovery post deshard until the recovery window expires.
  • WARNING: Replication delay intercept recovering – real value within range: Replication delay intercept is above or below the desired range, but the real value has already recovered. Replication delay intercept is trending towards recovery.
  • CRITICAL: Replication delay intercept is out of bounds, and slope does not indicate recovery: The replication delay is outside of the desired range, and is not moving towards recovery or is not moving towards recovery at the minimum desired rate (slope).
  • CRITICAL: Replication delay intercept recovery exceeded window: The replication delay intercept was recovering but exceeded the maximum recovery window time limit. Replication may be lagging or is not catching up in sufficient time.
  • WARNING: Active backup/deshard: An active backup or deshard job is running.

Naturally, other parameters necessary for healthy replication should continue to be checked as well. Replication, including delayed replication, will always fail if the replication state is stopped due to an error.

Caveats

No solution is completely perfect or without trade-offs, and delayed replication is no exception. Some considerations that should be taken into account prior to implementation:

  • If failing over to a delayed standby/slave, recovery time can be impacted. For instance, if auto-incrementing columns are used in MySQL and statement based or hybrid replication is being used then the standby must be caught up or data inconsistencies are likely. Make sure that the additional time to catch the host up is taken into account when estimating service restoration time.
  • The longer replication is delayed, the longer it will take for an instance to catch up.
  • If the application being targeted uses the delayed standby as a read slave, it is important to verify that the delay line doesn’t impact functionality or create edge/corner cases (this is not the case at IMVU, but is worth mentioning as it is not uncommon). Ideally, this should be validated by understanding the application’s logic or (even better) its code. A great example: A user is disabled in an application that actively queries read slaves for user information in a table. If the read slave is two hours behind, the user    might be able to log in until the slaves commit the transaction which disabled the account.
  • Delayed replication will not mitigate data loss unless the impacting event is caught within the delay window, and action is taken prior to the statements in question being executed. Clearly document that delayed replication only serves as a component of a comprehensive ecosystem.

Final Thoughts

Delayed replication is most useful to recover from errors made by administrative users, particularly since statements involving most DDL cannot be wrapped in a transaction. It is not, however, a magic bullet; it only provides protection if the offending transactions are identified and removed before making it through the delay line. The monitoring strategies for delayed replication are different (and decoupled) from standard MySQL replication monitoring, and the decoupled processes themselves must be monitored as well.

Continuous Monitoring: Real-time statistics for a thousand servers and the application they serve

By Jon Watte, Technical Director, IMVU.com 
Image

At IMVU, we push code to production fifty times a day. Each time an engineer finishes a task, the code goes through a large battery of unit tests, and when it passes, we deploy it on our servers right away. This makes the feedback loop immediate: If something is wrong, we hear about it and can fix it while the context is still fresh in the mind of the engineer.

An important part of this process is the “immune system.” The immune system monitors the status of the entire application, and detects abrupt changes. If these abrupt changes are bad enough, and closely enough correlated with a recent code deployment, that code deployment is rolled back, and the engineer in question sent links to graphs and error logs to go look at to figure out the problem.

For a long time, we used rrdtool with scripts to scrape counter values out of memcached to capture data, and cacti to plot that data into graphs. This was an easy way get get started when IMVU was small, and it has scaled to the size we’re at now. Two years ago, the system started showing its age. A year ago, we decided to do something about it. The problems we wanted to solve were:

1) The system we had only collected data at 5 minute intervals. This is way too slow to quickly detect problems after a bad code push. Bad code pushes are rare, but we want them to impact customers as briefly as possible.

2) The system we had would aggregate data as “average” over time, to keep coarser data available for a longer time. But this means that we lose useful resolution. What was the swing of the data within each “bucket” of measurements? What was the min, and the max?

3) The retention times for the data were too short. To compare if the system is mis-behaving right now, or if it’s just normal high load for a week-end, we need accurate data from a week ago as a baseline.

4) The system that relied on metrics to be written into memcache, and then scraped back out into rrd files by cacti, was running out of steam, and we often had time intervals with missing data for many counters.

To solve these problems, we went looking for other counter management solutions. We tried a large number, wrote off a bunch, and then settled on “Graphite,” which our friends over at Etsy seemed to recommend highly. However, Graphite was still not quite right — it would still only allow a single aggregation function when aggregating metrics over time, and the built-in storage back-end had some performance problems, largely traced to the distribution model of “use NFS.”

So, we started writing our own back-end for the nice Graphite graphing front-end. We made the back-end fit into Graphite’s expectations, and exposed the different data from a single metric as separate sub-counters. For each data point in a graph, we could get the average, sample count, standard deviation, minimum, and maximum. Getting there required pretty heroic efforts, and pretty nasty hacks, though — the internals of Graphite simply weren’t made to support this use case. Also, Graphite used server-side rendering, which meant that just a few engineers keeping a dashboard of a dozen counters on their screen, refreshing every 10 seconds, would overload the machine collecting the metrics.

At some point, enough was enough, and we took the back-end we’d developed, and wrote our own front-end. This front-end is a HTML5 application using client-side JavaScript for rendering, thus offloading the metrics server. It also uses HTTP for data transport, thus lending itself well to various kinds of clients — including web caching if needed!

Finally, to solve the intermittent data problem, we made the system capable of forwarding incoming data in a graph. An agent can run on each server, and receive local data, which it then forwards to the master database. Should the agent connection go down, the data is buffered while the agent attempts to re-connect.

Today, we’re releasing this (both back-end and front-end) on GitHub to the open source world as our contribution to operations and engineering teams everywhere. If you have a large number of counters to track, and want richer data than a “simple” aggregation function per data point, I encourage you to have a look, starting with the wiki:

https://github.com/imvu-open/istatd/wiki

The back-end application is written in C++ with boost::asio for threading and networking, and currently keeps half a million counter files, each updated every 10 seconds, on a mid-range Dell server with raid5 SSD drives. Currently, there is build and packaging support for Ubuntu 10.04 LTS, although any reasonable UNIX with GCC should be supported. Give it a spin, and let us know what you think!

Writing Resilient Unit Tests

By Andrew Wilcox

I happened to be looking at a unit test I wrote a couple months ago and was dismayed to realize it wasn’t working any longer:

test("don't launch the rocket in stormy weather", function () {
    setup_rocket_systems_controller();
    setup_fake_ground_based_internet();
    set_weather_to_stormy();
    equal(ready_to_launch(), false);
});

Today this test will always go green – regardless of whether the code launches the rocket in stormy weather or not.

My unit test worked when I first wrote it.  But my test was also fragile: it was easily broken in the normal course of code development.

Here’s the code being tested as it appears today:

function ready_to_launch() {
     var controller = connect_to_rocket_systems_controller();
     if (! controller) {
         return false;
     }

     if (controller.get_fuel_level() < needed_fuel()) {
         return false;
     }

     var internet = controller.connect_to_ground_based_internet();
     if (! internet) {
         return false;
     }

     if (internet.obtain_weather_report() == WEATHER_STORMY) {
         return false;
     }

     return true;
}

Of course we could argue whether this is the best coding style to use: whether each check could be split out into a separate function, or whether exceptions should be thrown instead of returning false, and so on.  But with working unit tests such improvements in design are easy to make.

Without unit tests (or with broken unit tests!) changing the code can easily break things without us noticing.  With my broken unit test, someone could be working on improving or enhancing the code, and they could make a small typo that broke the code which prevents launches in stormy weather, and they would think that everything was OK because — after all — there’s clearly a unit test that checks for that!

Why did my unit test break?  What happened was:

  • We noticed that launching our rockets without enough fuel often resulted in a crash;

  • An engineer carefully reviewed all the changes that would be needed to implement a fix, and wrote unit tests to check that the code would not launch the rocket when it didn’t have enough fuel;

  • A couple places in the code that needed to be updated were overlooked but caught by unit tests going red, and so were quickly found and fixed;

  • A few unit tests broke by going red because they themselves now needed to fuel the rocket in order to do their test, but they were also quickly found and fixed;

  • My unit test also now needed to fuel the rocket to do its test, but what it checked was just that the rocket wasn’t ready to launch… and so it was now going green not because of stormy weather but because it was failing to fuel the rocket!

At this point the code to check for stormy weather could be inadvertently broken by further code changes and no one might notice.  We could even delete the check for stormy weather entirely and my unit test would still go green!  Bit rot has set in… there’s impressive looking code to avoid launching the rocket when we’re not supposed to… and impressive looking unit tests to test that… and yet when stormy weather comes we could actually launch the rocket anyway.

There are two ways for a unit test to be broken, false-red tests and false-green tests:

  • false-red tests go red even if the code feature they’re testing is correct;

  • false-green tests go green even if the code they’re testing is incorrect.

Because we run all unit tests automatically before pushing to production, red tests are quickly found and fixed.  Thus any buggy unit tests that might be able to sneak into the code base are the false-green ones: the ones that say “fine!  everything’s fine!  no problem!”… even when the code is broken.

“Happy path” unit tests rarely break by going false-green.  A happy path test is one that checks that the desired result is achieved when all the necessary conditions are in place.  E.g., we launch our rocket, and it makes it to low Earth orbit, and it inserts into the correct orbit, and it matches velocity with the space station, and it docks without crashing, etc. etc.  Or, in the case of IMVU, a happy path test would be that customers can place orders when they have enough credits, the product is available, and so on.

Happy path unit tests rarely go false-green because of entropy: there are typically many more ways for something to fail than there are for something to succeed.  It’s possible that I might manage to break a test and break the code at the same time and have my changes conspire to get the test to go green anyway…

test("2 + 2", function () {
    equal(plus(2, 2), 5);
})

but it’s rather unlikely.

“Sad path” unit tests on the other hand check that things do in fact fail when they’re supposed to.  “The rocket can’t be launched in stormy weather” is one example.  Another example would be “customers can’t buy a virtual product when they don’t have enough credits”.

Sad path unit tests are often important for the integrity of the business.  Suppose the check for not being able to buy a product when the customer had insufficient credits wasn’t working.  We might not notice this — just using our application ourselves — because when we use our application we don’t generally try to buy things when we don’t have enough credits.

But though they are often quite important, these sad path unit tests are more easily broken as the code continues to be developed because there are many ways that things can fail — and if the test is just checking for “something failed” then lots of things might be able to do that.  For example, this unit test will go green if the code throws any exception:

test("names must be a string", function () {
    raises(function () { set_name(3.14159) });
});

and so it will be broken by any bug that manifests by throwing an exception.  In fact when I review existing unit tests that merely check if an exception has been thrown, I’ve found that they’ve become useless more often than not!

Thus extra care is needed when writing sad path tests.  It’s not enough to just check whether the code failed — we need to check whether it specifically failed for the reason that we’re testing for.

For this we need to do an analysis: are there any other code paths that could generate the same result as we’re testing for?  For example, it is better to check for the specific exception being thrown — but that could still be buggy if there’s more than one way that specific exception could be thrown.

In the rocket example, one option would be to instrument the code to make it more testable by having it return a string indicating why we’re not ready to launch:

function ready_to_launch() {
    var controller = connect_to_rocket_systems_controller();
    if (! controller) {
        return "the rocket systems controller is offline";
    }

    if (controller.get_fuel_level() < needed_fuel()) {
        return "there is insufficient fuel to launch";
    }

    var internet = controller.connect_to_ground_based_internet();
    if (! internet) {
        return "the ground based Internet service is not available";
    }

    if (internet.obtain_weather_report() == WEATHER_STORMY) {
        return "flight protocols do not permit launching in stormy weather";
    }

    return null;
}

Code will often need some enhancements to make it testable, though you may be able to find ways that make the testability requirement less intrusive.

And so if I had written my unit test at the time to check for the specific condition I was testing for:

test("don't launch the rocket in stormy weather", function () {
    ...
    set_weather_to_stormy();
    equal(
        ready_to_launch(),
        "flight protocols do not permit launching in stormy weather"
    );
});

then when the fuel requirement had been added later my test would have broken by going falsely red instead of going falsely green — and it would have been fixed.

My test would have been a more resilient unit test — one that over time would have had a better chance of being kept working as development proceeded.

Software engineering and product development best practices at IMVU

Follow

Get every new post delivered to your Inbox.

Join 53 other followers

%d bloggers like this: