A little over a year ago, Elon Musk acquired Twitter for $44 billion. A lot of PR turbulence followed, including a rebrand, ever-shifting revenue models, the platform’s API getting fenced off and free speech-related controversies. Regardless, there’s one decision that appears to have paid off in spades: the company’s shift away from the cloud and towards on-prem.
X’s Engineering team(@XEng) recently revealed all the changes that have taken place under the hood of the social media platform over the past few months. The developers have drastically downsized X’s codebase, with the personalized “For You” feed being slimmed down by a staggering 90 percent. As a result, the compute overhead has been cut in half. In addition, this feature now actually works better than before. Other software-based optimizations involve the API, tech stacks for a variety of features on X and a more effective blockade of content-scraping bots. This last initiative has upped the takedown count of such bots by 37 percent compared to 2022. DM spam has also reduced dramatically: there’s apparently 95 percent less of it now versus last year.
It will be music to the ears of Elon Musk, who hasn’t been shy to voice his displeasure with the giant bot population on X as well as other inefficiencies in the company. However, few Musk-related incidents in the past year have been as glaring on our end as the battle with cloud giant AWS: in June, X had failed to pay its bills and negotiations for a new cloud contract reached a dead end. The solution: much more on-premise in tandem with an equally vast reduction in the reliance on cloud services.
Cost savings
The X engineering team lists some impressive cost reductions as a result of the strategy switch. Monthly cloud costs decreased by 60 percent, thanks in part to the removal of media/blob artifacts from cloud environments. This alone reduced the cloud storage amount by 60 percent. Data processing costs were reduced even more drastically, leading to a savings of 75 percent.
To make the cloud exit a success, the company also optimized its own hardware stack. Thus, the company closed its Sacramento data center and reallocated the 5,200 racks and 148,000 servers inside, saving $100 million and making 48W available for other tasks in the process. Proprietary on-prem GPU Supercompute clusters were built, now relying on 43.2Tbps of network fabric architecture. The capacity of the company’s network backbone has been optimized as well, saving a hefty $13.9 million per year. Automated peak traffic failover tests continuously monitor platform scalability and availability. In short, it’s a laundry list of improvements and savings.
Independence
It seems that X has not only achieved more independence, but has massively reduced costs by utilizing its own hardware. Senior Director of Product Strategy at Veeam Rick Vanover recently (article in Dutch) shared how the cloud also leads to unforeseen costs for countless other parties. “We regularly hear stories about monthly bills for services that are 25 to 50 percent higher than the company had budgeted,” he said. While these are occasional inefficiencies that cause costs to skyrocket, fundamentally, cloud expenses arn’t as manageable and predictable as an on-prem solution.
X, of course, is of a size that the economies of scale works in the platform’s favour when building on-prem solutions. Nvidia CEO Jensen Huang, who as the leader of the largest GPU vendor by far plays a crucial role both in the cloud and for on-prem, said a few years ago “The more you buy, the more you save.” Therefore, it doesn’t make sense for everyone to choose on-prem over the cloud.
However, Vanover does emphasize the importance of a cloud exit strategy, which requires an organization to continually look at what cost savings are possible. What if it’s better to move to another cloud player or to deploy a local data center? The flexibility to withdraw from the chosen cloud environment should be guaranteed even before choosing the cloud service, according to Vanover. “It’s like when you’re a firefighter or police officer: the first thing you’d think about when entering a potentially dangerous situation is how to get out, if necessary.”
In other words: X benefits from this “cloud exit,” but for other parties, the outlook may be different. Those who purchase hardware are going to need to ensure that it’ll bring a return on investment. Also, upgrades can be prohibitively expensive, especially if there’s a sudden hype around specific hardware as we’ve seen in 2023. On-prem is therefore not immune from the shifting sands of supply and demand. Still, X’s example shows that, as a company (regardless of size), you need to stay flexible enough to shift your IT infrastructure away from the cloud, should it be beneficial to do so.
Also read: NetApp makes all-flash and AI accessible to more organizations
2 days ago
