links for 2010-03-05
March 5, 2010
-
Finally. Blocks flash from loading unless you click it first.
Real-Time is the new Web2.0: Both mean nothing.
March 5, 2010
So you come across a startup that’s pitching a “Real-Time” service, what do you do? Punch them in the face – now that’s real-time! Well, maybe that’s a bad idea, but you should completely tell them they’re not getting anywhere just by calling their service “Real-Time”. Here are some examples of concepts which can’t ever be real-time:
1. The News – Maybe if people as a whole were more intelligent, but, the closest you’re going to get is Digg or Reddit and those require thousands of data points (over time) before an article bubbles to the top of the relevancy list. The exception of course is “Bad News”, that could easily be done in real-time.
2. Product Pricing - Retailers have a hard enough time with loss prevention and maintaining profits than to care if their published prices and inventory are accurate or not. Sure, they have real-time inventory internally, but that’s a large enough dataset that it’ll never be replicated to a service provider; the short story is that you’ll never be able to get both an instant price and instant data at the same time.
Maybe if you were Google.
3. Search (sites, news, or otherwise) – Indexing is hard and there’s only one cat in the game with the facilities to do so in real-time. The only problem is the rest of the world doesn’t have a supercomputer running their system and there will always be a delay before Google gets the memo. The exceptions here are sites that Google cares about but chances are you’re not going to be big enough for that, else, you probably wouldn’t be a startup.
4. Communications – There’s already an “app for that”. It’s called the phone and your voice. Pickup phone, call friend, profit. Anything else might provide real-time delivery on one end or the other, but chances are, one person in the party is playing a video game, watching youtube, or chatting on Facebook in which case their response will be in Internet time.
Let me just steal the definition of Real-Time from Wikipedia:
In computer science, real-time computing (RTC), or “reactive computing”, is the study of hardware and software systems that are subject to a “real-time constraint”—i.e., operational deadlines from event to system response. By contrast, a non-real-time system is one for which there is no deadline, even if fast response or high performance is desired or preferred. The needs of real-time software are often addressed in the context of real-time operating systems, and synchronous programming languages, which provide frameworks on which to build real-time application software.
If that’s not your product then please, stop calling yourself real-time and get an old comp sci book, then figure out what real-time really is.
Don’t steal your users’ domains (wordpress.com)
March 4, 2010
This is primarily with regard to WordPress.com, but, it’s an important and delicate concern for any service. Currently, if you want yourdomain.com to point to your WP blog, you have to give WP control over your domain name. WordPress has the challenging problem of providing scalable content and this is very hard to do while providing domain name support at the same time; this is why WP requires control of your domain’s DNS. WordPress has a highly redundant and robust network and part of that is having control over their user’s domains. This is limiting in that a user can’t point a subdomain like securepayments.yourdomain.com to another service.
While DNS natively supports this functionality, it’s not something the average blogger, hosting provider, or even application developer understands (it’s called Slaving). If/when WordPress supports this functionality, they’ll need to do so carefully to limit exposing their systems to outside risks, make it easy to use, and support the most complex users who need to run their own DNS.
If you provide a web service, such as a blogging platform, don’t build “domain name support” by forcing your users to give you full control over their domain name. In short, I don’t blame WP for their current implementation (it’s probably a lower priority for them), but here’s a few tips on what you should do if you want to make YOUR app/service support user-supplied domain names:
- Don’t give the feature away for free. Your support costs will be higher for this feature than most.
- Don’t give the feature away for free. DNS plays a big role in spam and you don’t want to associate your systems with a spammer’s domain name.
- Think very hard if you want to point a user’s app to the root of their domain. Doing so opens you up to the concerns above.
- If your service involves high-traffic content such that load balancers are involved, consult someone who’s figured this problem out before. (my advice is typically free)
- Whenever you’re building and scaling your services, keep all the above in mind. As the Internet becomes connected at the application layer, via things like federations and digital certificates, domain names will become more and more important as a component of security and authority.
- Let users host their own DNS and point it at you instead of the other way around. Administering DNS servers isn’t very cost effective and you’ll save money leaving it up to the users. The downside is you’ll need to be very good at providing your users with instructions and you’ll need to notify them when changes are necessary; again, Don’t give the feature away for free.
- Let users host their own DNS. You don’t want to host DNS for your users — doing so is a long-term cost commitment and something you can’t “undo”.
links for 2010-02-13
February 13, 2010
links for 2009-11-26
November 26, 2009
Amazon RDS: Poison or Pill
October 29, 2009
As soon as read the AWS newsletter about Amazon RDS, I started looking for a Megaphone to start shouting at folks – keep away! Amazon RDS or Relational Database Service places Amazon into the mire of shared hosting and AW users into a position of false confidence. Harsh words considering, overall, I feel Amazon’s service offerings are best-in-class. AWS offerings have historically pushed the envelope with regard to practical usage-based computing, something which ancient providers such as Sun and IBM have attempted to accomplish for decades; in this case I define practical as both usable and cost effective for small and large tasks. Up until now such systems weren’t trivialized to x86 hardware and required special programming considerations, access to academic institutions and/or a large budget. By combining SLA-supported x86 virtualization alongside application services such as S3, SQS, and SimpleDB, AWS has provided a usage-based on-demand computing solution which is simpler than task-based computing and as secure and reliable as virtualized or shared hosting. With it’s on-demand nature AWS is a cost effective for everything from small tasks to those requiring a datacenter of processors.
So why is Amazon RDS so bad, so much that you shouldn’t use it?
Well, there’s not an easy answer, the better question is to ask yourself: Why do you think AWS will be better than your own MySQL deployment? There is no right answer because almost any answer will probably, one day, bite you in the ass. Hard. I mean data loss, and it won’t be Amazon’s fault.
RDBMS systems and applications which depend on them are built from the ground up to rely on persistence, integrity, and static data models (schema). In contrast AWS has been built for distribution, decentralization, and the “cloud”. For Amazon, this service is somewhat of a U-turn from their original direction and has also placed a stamp on their forehead which says “That MySQL Guy” which is not good — I have nothing against mysql, however, as a de facto entry-level (free open source) software, it has accrued a strong following of immature software. Such software has nothing to do with the basic purposes of AWS or MySQL but has everything to do with how Amazon’s support and engineering staff will be spending their time which is supporting users and software which aren’t built for the cloud.
I hope that RDS won’t be a situation of butterflies & hurricanes but here’s a quick list of why the relative cost of RDS is high both for Amazon (the company) and all of it’s AWS users:
- Cost for Amazon (operations, engineers, and products)
- MySQL, like most open source systems, has been historically buggy software with a trailing release+testing+production schedule which requires continuous testing between production releases for large deployments (such as RDS).
- MySQL has a large set of features which vary across releases and which share equal presence in production; in other words, Amazon will need to cater to providing production support for multiple versions, not just the latest stable version.
- Amazon has no control over features and capabilities of MySQL and is thus limited to what MySQL provides; while MySQL provides many “good things”, Amazon will still be obligated to maintain through the bad. This is a shared disadvantage of AWS Map Reduce via Hadoop however, those are mostly mitigated because Map Reduce is such a low-level distributed system.
- MySQL is very flexible and itself scales very well however it doesn’t do so by itself and requires a significant effort to be properly configured for the data being managed. All the folks who don’t know this will default into thinking Amazon will do this for them and will be disappointed when it doesn’t “just work”. Whether they ditch RDS or bug Amazon’s support, either way, it’s not a positive situation.
- Cost for AWS (primarily EC2) users
- Potential degradation of service and support for EC2 instances
- With RDS available Amazon can defer issues with regard to running MySQL on EC2 instances to a recommendation for RDS — this will be a terrible waste of time for both parties.
- MySQL is a very centralized system and by transitioning the decision of where MySQL resides in the AWS cloud from the user to Amazon, Amazon will be further centralizing the impact of MySQL on the cloud. Whereas users will randomly have MySQL deployed across any EC2 instance, Amazon will be appointing MySQL to specific hardware; this is based on the assumption that Amazon is clustering RDS deployments onto local hardware and not randomly deploying instances in the cloud. This is somewhat of a compromise for security and adds significant SLA risks (read: cost) to Amazon. In short, when a MySQL cluster dies – a LOT of folks are going to be VERY unhappy – their support tickets will be a burden to staff and their requests for credits will be a financial cost. Moreover, support staff will be yielding priority to these customers over other services because of the implicit severity.
- Increased cost
- RDS instances cost >10% more than regular instances and only come with the added benefit of backups — something which every system should already have in place. If you do choose to delegate the task of backups to RDS, you’re paying extra for a task you’ve already thought about doing yourself.
- Cost of keeping your database, it’s backups, and it’s history all within AWS is multiplicative and if you grow to the point where you’re ready to move off you’ll be charged to transfer all the data to an external system. While this is a subjective cost it’s still worth pointing out; if folks aren’t already doing backups right, they’ll likely not know that cost effective database backups make use of binary logging facilities, not filesystem snapshots, and use significantly less disk space (and thus I/O).
- False confidence
- As I’ve mentioned before, letting other folks control your backups for you is a mistake. Failure is a matter of when, not if, and you’ll be in better control of responding if you understand what you’re dealing with. Just because RDS is doing you’re backups doesn’t mean you’re safe.
- RDS users will expect MySQL to scale on-demand as everything else works that way with AWS and it’s just not that simple. Scaling a database requires analysis and a balanced combination of server settings, data normalization, and indexes; all of these things will still be the user’s responsibility and Amazon’s solution of “throw hardware at it” is a haunted path to send it’s users down.
- Potential degradation of service and support for EC2 instances
Overall, I feel that Amazon could quickly cannibalize the value and quality of AWS if they (continue to) introduce trivial services. Supporting open source software they have no control over is a significant increase in relative support and operations cost. Amazon seems to be approaching this by making the cost of RDS instances more than EC2 which is a mistake because the real cost is the lost opportunity of engineers spending their time on systems which are more efficient for cloud computing – Amazon could charge 3 times an EC2 instance and their engineers would still be better off building technologies for cloud-based systems and not centralized RDBMS-dependent web applications.
Where I feel Amazon has fallen short the most, is that RDS only provides single-instance MySQL support and nothing more. No load balancing, replication, Hadoop integration, or any other form of data abstraction which could make it functional in a cloud computing context. Not implementing these features is a very clear indicator that AWS is focused more on short term revenue generating feature rather than cost effective cloud computing systems or improving the shortfalls of legacy centralized system.
With all this said, I have to consider the possibility of this being a good move for Amazon. I present the potential issues with RDS simply to warn folks from relying on it as a crutch, and, to point out the new direction AWS has veered is into choppy waters. There are several aspects of RDS which will give Amazon insight into correlations among and between the varying systems of data storage and processing – comparing SimpleDB, MapReduce, MySQL, and general resource consumption could shed light onto how their cloud is being used at a higher level than processors and bandwidth. Last, Amazon might be aware that MySQL is a crutch and is putting the service out there as a way to wean folks off of centralized systems.
links for 2009-10-22
October 22, 2009
Ruby EventMachine :gt Python Tornado, Twisted
September 25, 2009
I was working on a project looking to see if Ruby was good enough for responding quickly to HTTP requests. Good thing it, along with Python, and every other language, plays well with C/C++. Anyways, EventMachine apparently blows away Tornado and Twisted. I only tested Tornado because it’s faster, right? What I really wanted to test was if either of these would fall apart under high concurrency or load. For the “Hello World!”, they both survived although as you can see for Tornado, response times became an issue earlier. I’ve also provided ‘ab’ for reference – it’s a little more specific with regard to response times. Clearly both of these are hitting a CPU ceiling – with Tornado hitting it faster. Ftr, I tested on a dual-core 2.33ghz xeon w/RHEL5, python2.6, and ruby1.8.5.
Along my adventure in this hnews thread, I came along this most awesome post: Twisted vs. Tornado: You’re Both Idiots
Anyways, what I’m happy about is there’s a Ruby option for a fast little server which pumps out a bajillion requests per second if you’ve got a farm of servers and it won’t fall on it’s face. Also, I don’t have to use Python and EventMachine is a BREEZE to use. What does suck is the EM HTTP server isn’t RFC compliant but that’s probably just a matter of time and I won’t be using HTTP anyways. ymmv
httperf: Tornado
[root@mail ~]# httperf --port=3002 --num-conns=1000 --num-calls=500 --rate 100 -v httperf --verbose --client=0/1 --server=localhost --port=3002 --uri=/ --rate=100 --send-buffer=4096 --recv-buffer=16384 --num-conns=1000 --num-calls=500 httperf: maximum number of open descriptors = 1024 reply-rate = 5045.8 reply-rate = 4868.5 reply-rate = 4905.4 reply-rate = 4846.9 reply-rate = 4938.4 reply-rate = 4747.3 reply-rate = 4800.2 reply-rate = 4795.6 reply-rate = 4595.3 reply-rate = 4591.1 reply-rate = 4784.6 reply-rate = 4775.9 reply-rate = 4563.3 reply-rate = 4872.3 reply-rate = 4948.8 reply-rate = 4853.0 reply-rate = 4551.3 reply-rate = 4587.3 reply-rate = 4885.7 reply-rate = 4900.2 Maximum connect burst length: 1
Total: connections 1000 requests 500000 replies 500000 test-duration 104.059 s
Connection rate: 9.6 conn/s (104.1 ms/conn, <=1000 concurrent connections) Connection time [ms]: min 34704.2 avg 93867.4 max 97177.5 median 95862.5 stddev 6293.6 Connection time [ms]: connect 0.0 Connection length [replies/conn]: 500.000
Request rate: 4805.0 req/s (0.2 ms/req) Request size [B]: 62.0
Reply rate [replies/s]: min 4551.3 avg 4792.8 max 5045.8 stddev 144.4 (20 samples) Reply time [ms]: response 187.7 transfer 0.0 Reply size [B]: header 156.0 content 12.0 footer 0.0 (total 168.0) Reply status: 1xx=0 2xx=500000 3xx=0 4xx=0 5xx=0
CPU time [s]: user 2.97 system 99.60 (user 2.9% system 95.7% total 98.6%) Net I/O: 1079.2 KB/s (8.8*10^6 bps)
Errors: total 0 client-timo 0 socket-timo 0 connrefused 0 connreset 0 Errors: fd-unavail 0 addrunavail 0 ftab-full 0 other 0
httperf: Ruby EventMachine
[root@mail ~]# httperf --port=3001 --num-conns=1000 --num-calls=500 --rate 100 -v httperf --verbose --client=0/1 --server=localhost --port=3001 --uri=/ --rate=100 --send-buffer=4096 --recv-buffer=16384 --num-conns=1000 --num-calls=500 httperf: maximum number of open descriptors = 1024 reply-rate = 11631.7 reply-rate = 9769.5 reply-rate = 9352.3 reply-rate = 10086.1 reply-rate = 8899.4 reply-rate = 9759.3 reply-rate = 9985.1 reply-rate = 10152.8 reply-rate = 10383.9 Maximum connect burst length: 1
Total: connections 1000 requests 500000 replies 500000 test-duration 49.590 s
Connection rate: 20.2 conn/s (49.6 ms/conn, <=984 concurrent connections) Connection time [ms]: min 229.8 avg 39130.7 max 42870.4 median 41409.5 stddev 6775.5 Connection time [ms]: connect 0.0 Connection length [replies/conn]: 500.000
Request rate: 10082.7 req/s (0.1 ms/req) Request size [B]: 62.0
Reply rate [replies/s]: min 8899.4 avg 10002.2 max 11631.7 stddev 756.9 (9 samples) Reply time [ms]: response 78.3 transfer 0.0 Reply size [B]: header 65.0 content 12.0 footer 0.0 (total 77.0) Reply status: 1xx=0 2xx=500000 3xx=0 4xx=0 5xx=0
CPU time [s]: user 2.20 system 46.84 (user 4.4% system 94.4% total 98.9%) Net I/O: 1368.7 KB/s (11.2*10^6 bps)
Errors: total 0 client-timo 0 socket-timo 0 connrefused 0 connreset 0 Errors: fd-unavail 0 addrunavail 0 ftab-full 0 other 0
ab: Tornado
[root@mail ~]# ab -c1000 -n100000 http://127.0.0.1:3002/
This is ApacheBench, Version 2.0.40-dev <$Revision: 1.146 $> apache-2.0
Copyright 1996 Adam Twiss, Zeus Technology Ltd, http://www.zeustech.net/
Copyright 2006 The Apache Software Foundation, http://www.apache.org/
Benchmarking 127.0.0.1 (be patient)
Completed 10000 requests
Completed 20000 requests
Completed 30000 requests
Completed 40000 requests
Completed 50000 requests
Completed 60000 requests
Completed 70000 requests
Completed 80000 requests
Completed 90000 requests
Finished 100000 requests
Server Software: TornadoServer/0.1
Server Hostname: 127.0.0.1
Server Port: 3002
Document Path: /
Document Length: 12 bytes
Concurrency Level: 1000
Time taken for tests: 27.996766 seconds
Complete requests: 100000
Failed requests: 0
Write errors: 0
Total transferred: 16800336 bytes
HTML transferred: 1200024 bytes
Requests per second: 3571.84 [#/sec] (mean)
Time per request: 279.968 [ms] (mean)
Time per request: 0.280 [ms] (mean, across all concurrent requests)
Transfer rate: 586.00 [Kbytes/sec] received
Connection Times (ms)
min mean[+/-sd] median max
Connect: 0 197 1102.8 0 20998
Processing: 1 50 37.3 45 5234
Waiting: 0 49 37.4 44 5234
Total: 18 247 1109.2 45 21253
Percentage of the requests served within a certain time (ms)
50% 45
66% 48
75% 52
80% 57
90% 77
95% 1237
98% 3074
99% 3112
100% 21253 (longest request)
ab: EventMachine
[root@mail ~]# ab -c1000 -n100000 http://127.0.0.1:3001/
This is ApacheBench, Version 2.0.40-dev <$Revision: 1.146 $> apache-2.0
Copyright 1996 Adam Twiss, Zeus Technology Ltd, http://www.zeustech.net/
Copyright 2006 The Apache Software Foundation, http://www.apache.org/
Benchmarking 127.0.0.1 (be patient)
Completed 10000 requests
Completed 20000 requests
Completed 30000 requests
Completed 40000 requests
Completed 50000 requests
Completed 60000 requests
Completed 70000 requests
Completed 80000 requests
Completed 90000 requests
Finished 100000 requests
Server Software:
Server Hostname: 127.0.0.1
Server Port: 3001
Document Path: /
Document Length: 12 bytes
Concurrency Level: 1000
Time taken for tests: 15.238117 seconds
Complete requests: 100000
Failed requests: 0
Write errors: 0
Total transferred: 7700077 bytes
HTML transferred: 1200012 bytes
Requests per second: 6562.49 [#/sec] (mean)
Time per request: 152.381 [ms] (mean)
Time per request: 0.152 [ms] (mean, across all concurrent requests)
Transfer rate: 493.43 [Kbytes/sec] received
Connection Times (ms)
min mean[+/-sd] median max
Connect: 0 76 603.3 0 9000
Processing: 0 32 264.1 15 10627
Waiting: 0 31 264.1 14 10625
Total: 9 108 752.8 15 14642
Percentage of the requests served within a certain time (ms)
50% 15
66% 15
75% 15
80% 22
90% 33
95% 35
98% 2999
99% 3015
100% 14642 (longest request)
links for 2009-09-07
September 7, 2009
-
Computer scientists are fond of talking about metadata. There often seems to be an assumption that drawing a distinction between metadata and data is useful and perhaps even necessary…
