Wikia's Bergman Urges SSD Adoption to Eliminate Database Sharding

O'Reilly

Flash storage removes the random I/O bottlenecks that slow down relational joins, allowing engineering teams with under 50 terabytes of data to bypass complex sharding architectures entirely with an $80,000 server.

AI-assisted summary of Velocity 2011: Artur Bergman, "Artur on SSD's" — original source. Check the original for full context.

Source published: 2011-06-20

Key ideas

Artur Bergman argues that mechanical hard drives cause massive productivity loss, noting a boot time of 12 seconds on SSD versus 60 seconds on spinning disks. (0:02)
When measured by gigabytes per IOPS rather than raw storage capacity, solid-state drives are significantly cheaper than traditional spinning disks. (1:16)
A single server equipped with 50 terabytes of SSD capacity for around $80,000 can eliminate the architectural complexity of database sharding for most workloads. (1:16)
SSDs cut drive-level power consumption from roughly 15 watts down to 1 watt while delivering 44,000 IOPS compared to just 180 IOPS on a mechanical drive. (1:16)

Detailed notes

## Background Published by O'Reilly on YouTube on June 20, 2011, this Velocity 2011 lightning talk features Artur Bergman, then VP of Engineering and Operations at Wikia and founder of Fastly. The video description highlights Bergman's extensive infrastructure background—ranging from high-volume trading architectures and LiveJournal engineering to federated systems—as he delivers a fast-paced case for solid-state drive adoption.

## TL;DR Bergman challenges the tech industry's hesitance to adopt SSDs, arguing that failing to equip developer laptops and data centers with flash storage wastes immense human and financial capital. He demonstrates that boot-time productivity gains alone justify laptop upgrades, while server-grade SSDs completely outclass mechanical hard drives on cost per IOPS, throughput, and power consumption. By deploying high-density commodity flash storage, engineering teams can achieve gigabytes-per-second random write performance and eliminate the architectural headache of database sharding for datasets up to 50 terabytes.

## Atmosphere & style The talk carries the irreverent, high-octane energy of a classic tech conference lightning pitch, driven by Bergman's blunt, unfiltered delivery and charismatic frustration. He commands the room with deadpan provocation—bluntly telling attendees they are wasting their lives on spinning disks—while backing his exasperation with rapid-fire back-of-the-napkin math and hardware benchmarks. The vibe is punchy, practical, and entertaining, ending on a humorous high note as the host jokingly dubs him the Kanye West of storage tech.

## Bulletin - [00:02](https://www.youtube.com/watch?v=H7PJ1oeEyGg&t=2s) Artur Bergman argues that mechanical hard drives cause massive productivity loss, noting a boot time of 12 seconds on SSD versus 60 seconds on spinning disks. - [01:16](https://www.youtube.com/watch?v=H7PJ1oeEyGg&t=76s) When measured by gigabytes per IOPS rather than raw storage capacity, solid-state drives are significantly cheaper than traditional spinning disks. - [01:16](https://www.youtube.com/watch?v=H7PJ1oeEyGg&t=76s) A single server equipped with 50 terabytes of SSD capacity for around $80,000 can eliminate the architectural complexity of database sharding for most workloads. - [01:16](https://www.youtube.com/watch?v=H7PJ1oeEyGg&t=76s) SSDs cut drive-level power consumption from roughly 15 watts down to 1 watt while delivering 44,000 IOPS compared to just 180 IOPS on a mechanical drive.

## Key takeaways - [00:02](https://www.youtube.com/watch?v=H7PJ1oeEyGg&t=2s) Very few engineering organizations in 2011 were running their entire data center storage infrastructure on solid-state drives. - [00:02](https://www.youtube.com/watch?v=H7PJ1oeEyGg&t=2s) Mechanical laptop drives waste substantial developer hours during reboot cycles, easily justifying an estimated $500 SSD upgrade cost. - [00:02](https://www.youtube.com/watch?v=H7PJ1oeEyGg&t=2s) Across a 1,700-person conference crowd, standard one-minute reboot times waste an entire cumulative day of human productivity per reboot. - [01:16](https://www.youtube.com/watch?v=H7PJ1oeEyGg&t=76s) Engineering teams mistakenly avoid SSDs due to perceived upfront drive costs without evaluating the real cost metric: gigabytes per IOPS. - [01:16](https://www.youtube.com/watch?v=H7PJ1oeEyGg&t=76s) Database joins are not inherently bad; poor relational query performance is typically caused by slow random disk I/O on mechanical drives. - [01:16](https://www.youtube.com/watch?v=H7PJ1oeEyGg&t=76s) Wikia's SSD-based file server achieved traversal of 8 million files on NFS in 9 minutes and a full backup in 12 minutes. - [01:16](https://www.youtube.com/watch?v=H7PJ1oeEyGg&t=76s) Wikia's storage setup demonstrated sustained throughput of 2.2 gigabytes per second of pure random writes. - [01:16](https://www.youtube.com/watch?v=H7PJ1oeEyGg&t=76s) A single standard SSD delivers roughly 44,000 IOPS compared to 180 IOPS from a mechanical disk, effectively replacing around 50 spinning drives. - [01:16](https://www.youtube.com/watch?v=H7PJ1oeEyGg&t=76s) Flash storage drives consume approximately 1 watt compared to 15 watts for mechanical spinning disks. - [01:16](https://www.youtube.com/watch?v=H7PJ1oeEyGg&t=76s) Most companies with under 50 terabytes of data can avoid complex database sharding entirely by consolidating onto a single dense SSD server for about $80,000. - [01:16](https://www.youtube.com/watch?v=H7PJ1oeEyGg&t=76s) Organizations should adopt mainstream commodity SSDs, such as the Intel 320 series, rather than overspending on ultra-expensive niche enterprise cards. - [01:16](https://www.youtube.com/watch?v=H7PJ1oeEyGg&t=76s) Wikia and Fastly adopted laptops equipped with SSDs as standard signing gifts for onboarding new engineering hires.

## Detailed walkthrough ### [00:02](https://www.youtube.com/watch?v=H7PJ1oeEyGg&t=2s) Stop Wasting Your Life: Put SSDs in Laptops

Artur Bergman starts his lightning talk by taking a quick show of hands across the audience to gauge SSD adoption. While several attendees use solid-state drives in their personal laptops, fewer raise their hands for using them in data centers. When he asks how many people are running their data centers completely on SSDs—excluding backups and near-line storage—only a single hand goes up. Artur points out that he personally built that particular deployment. Looking out at the audience, he delivers a blunt assessment: “you're basically wasting your life like everyone who doesn't have an SSD in your laptop you're wasting your life”.

To prove his point, Artur showcases his Mac's boot screen, which takes only 12 seconds to start up. He does quick math on the conference crowd: with roughly 1,700 people attending Velocity, standard mechanical hard drives taking about a full minute to boot mean that every time the room reboots their computers, an entire day's worth of collective human time is thrown away. He urges anyone running engineering organizations to simply spend the $500 to equip developers' laptops with SSDs immediately, emphasizing that the productivity gains pay for themselves right away.

- Mac boot time with SSD: 12 seconds versus ~60 seconds on a standard mechanical hard drive. - Scale calculation: Across 1,700 attendees, a single standard reboot cycle burns an entire day of wasted productivity. - Estimated upgrade cost: Approximately $500 per laptop to immediately recoup lost developer time.

[01:16](https://www.youtube.com/watch?v=H7PJ1oeEyGg&t=76s) The Real Math: IOPS, Database Performance, and Eliminating Sharding

Artur vents his frustration that despite preaching the benefits of SSDs for two and a half years, engineers repeatedly give him the excuse that flash storage is too expensive. In reality, he argues, solid-state drives are significantly cheaper than spinning disks when evaluated on the metric that actually matters: gigabytes per IOPS. Traditional disk arrays fail because of disk I/O bottlenecks during random access patterns, not because relational database operations like joins are inherently flawed. He laments the circular reasoning common in database architecture: people claim joins are bad because they are slow, so they resort to sharding to avoid them, when the actual villain is slow disk I/O.

To demonstrate the real-world power of flash storage, Artur details Wikia’s new file server setup. The machine can traverse 8 million files on NFS in 9 minutes, perform a full backup in 12 minutes, and push 2.2 gigabytes per second of random writes. Because a single modern server can be packed with 50 terabytes of SSD capacity for roughly $80,000, Artur argues that most teams can bypass complex sharding schemes entirely: “you can get 50 terabytes of ssds in one machine for like $80,000 and you can fucking skip sharding unless you have more than 50 terabytes which is actually more most of you you know like just not needed”.

```mermaid flowchart TD accTitle: Database Scaling Decision Path accDescr: A comparison showing how SSD performance removes the necessity of architectural sharding for datasets under 50 terabytes.

subgraph Solution["The SSD Architecture"] E["Single Commodity SSD (44,000 IOPS)"] -->|"Delivers 2.2 GB/s Random Writes"| F["Fast Joins at Scale"] F -->|"Supports up to 50 TB Storage for ~$80k"| G["Skip Sharding Entirely"] end ```

Beyond eliminating software complexity, Artur highlights drastic efficiency and power savings: - Wikia file server performance: Handles 8 million files on NFS in 9 minutes and completes a full backup in 12 minutes. - Throughput: Sustains 2.2 GB/s of pure random write throughput. - Performance disparity: A single mechanical hard drive provides roughly 180 IOPS, whereas a single standard SSD can output around 44,000 IOPS, effectively replacing 50 spinning drives. - Power consumption: Flash storage drops power down to roughly 1 watt per drive unit compared to 15 watts for spinning disks. - Deployment advice: Start simple with mainstream commodity drives like the Intel 320 series (around $1,000 for 600 GB) instead of spending tens of thousands on niche SLC enterprise cards. As Artur quips, there is no need to jump straight to an unattainable Formula 1 race car when you are currently riding a bicycle—a "Ferrari" like a basic Intel SSD is more than enough.

Artur concludes by mentioning that SSDs serve as the default signing bonus gift for new hires joining Wikia (and Fastly). The session host wraps up by jokingly introducing him as "Artur Kanye West Bergman" with the closing riff: “I'mma let you finish but SSDs are the coolest disks of all time.”

## Notable quotes > [01:16](https://www.youtube.com/watch?v=H7PJ1oeEyGg&t=76s) you can get 50 terabytes of ssds in one machine for like $80,000 and you can fucking skip sharding unless you have more than 50 terabytes — Artur Bergman

> [00:02](https://www.youtube.com/watch?v=H7PJ1oeEyGg&t=2s) you're basically wasting your life like everyone who doesn't have an SSD in your laptop you're wasting your life — Artur Bergman

## Mentioned Artur Bergman, Wikia, Fastly, Velocity, O'Reilly, Intel 320 series, Mac, Kanye West, OpenID Foundation, LiveJournal, Six Apart, Perl 5, djabberd

## Video description Artur Bergman (Wikia/Fastly), Artur Bergman, hacker and technologist at-large, is the VP of Engineering and Operations at Wikia. He provides the technical backbone necessary for Wikia's mission to compile and index the world's knowledge. He is also an enthusiastic apologist for federated identity and a board member of the OpenID Foundation. In past lives, he's built high volume financial trading systems, re-implemented Perl 5's threading system, wrote djabberd, managed LiveJournal's engineering team, and served as operations architect at Six Apart. His current interests extend to encompass semantic search, large scale infrastructure, open source development, federated instant messaging, neurotransmitters, and the future of cyborgs.

[Source](https://www.youtube.com/watch?v=H7PJ1oeEyGg)

中文简报

Wikia 的 Bergman 呼吁采用 SSD 以消除数据库分片闪存存储消除了减缓关系联接的随机 I/O 瓶颈，使数据量低于 50 TB 的工程团队仅凭一台价值 8 万美元的服务器即可彻底避开复杂的分片架构。 Bergman 对科技行业在采用 SSD 上的犹豫态度提出质疑，认为未在开发者笔记本和数据中心配备闪存存储浪费了巨大的人力与财务资本。他指出，仅开机速度带来的生产力提升就足以证明升级笔记本的合理性，而服务器级 SSD 在每 IOPS 成本、吞吐量和功耗方面完全超越了机械硬盘。通过部署高密度商用闪存，工程团队可以实现每秒数吉字节的随机写入性能，并彻底消除 50 TB 以内数据集在数据库分片上的架构难题。 Artur Bergman 认为机械硬盘会导致巨大的生产力损失，并指出 SSD 的开机时间仅需 12 秒，而旋转磁盘则需要 60 秒。 (0:02) 如果以每 IOPS 对应容量而非原始存储容量来衡量，固态硬盘比传统机械硬盘要便宜得多。 (1:16) 一台配备 50 TB SSD 容量、成本约 8 万美元的单台服务器，即可消除大多数工作负载中数据库分片的架构复杂性。 (1:16) SSD 将驱动器级别的功耗从约 15 瓦降至 1 瓦，同时可提供 44,000 IOPS，相比之下机械硬盘仅有 180 IOPS。 (1:16)

Reading summary

Artur Bergman argues that the tech industry's reluctance to adopt solid-state drives stems from flawed cost assessments, asserting that flash storage dramatically improves developer productivity and eliminates the architectural need for database sharding. Developer Productivity and Laptop Upgrades (0:02) Bergman contends that engineering organizations waste substantial time by failing to equip developer laptops with solid-state drives. Demonstrating a 12-second Mac boot time compared to roughly 60 seconds on a conventional mechanical drive, he calculates that across 1,700 conference attendees, a single reboot cycle burns an entire day of cumulative human productivity. At an estimated upgrade cost of $500 per machine, Bergman maintains that the productivity gains immediately justify the expense. Resolving Database Bottlenecks Without Sharding (1:16) Bergman challenges the common architectural assumption that relational database operations, such as joins, are inherently flawed. He argues that queries become slow due to mechanical disk I/O bottlenecks under random access patterns rather than relational logic, leading teams to build unnecessary and complex sharding architectures. Because a single server can house 50 terabytes of SSD capacity for approximately $80,000, he asserts that organizations with datasets below that threshold can avoid sharding entirely. Storage Economics, IOPS, and Operational Efficiency (1:16) Evaluating storage on a metric of gigabytes per IOPS rather than raw capacity makes flash storage cheaper than spinning disks. While a typical mechanical drive yields about 180 IOPS and consumes roughly 15 watts, a standard SSD delivers approximately 44,000 IOPS while drawing around 1 watt, effectively replacing 50 mechanical drives. To illustrate real-world throughput, Bergman cites a Wikia file server installation that traverses 8 million NFS files in 9 minutes, executes a full backup in 12 minutes, and sustains 2.2 GB/s of random writes. He advises organizations to deploy mainstream commodity flash, such as the Intel 320 series at roughly $1,000 for 600 GB, instead of overspending on specialized enterprise SLC hardware. Bergman positions commodity solid-state drives as a decisive operational upgrade for both client laptops and server fleets, though his premise that teams can bypass sharding depends on keeping active storage requirements under the 50-terabyte single-server ceiling.

精炼阅读摘要

Artur Bergman在Velocity 2011大会上指出，机械硬盘造成的随机I/O瓶颈是导致数据库变慢与架构复杂化的根源；企业若换用固态硬盘（SSD），不仅能通过缩短开机时间提升员工效率，更能以极高性价比消除分片（sharding）等繁杂设计。个人设备升级对生产力的立竿见影提升 (0:02) Bergman以现场观众的笔记本使用情况为切入点，指出机械硬盘正在无形中消耗大量人力成本。他展示了搭载SSD的Mac电脑仅需12秒即可完成开机，而传统机械硬盘普遍需要约1分钟。按照Velocity现场约1,700名与会者的规模计算，如果全员使用机械硬盘，仅一次集体重启就会白白损耗整整一天的群体生产力。他主张技术团队主管应立即为开发者个人电脑配置SSD，尽管每台升级成本约为500美元，但由此换取的工时效率提升能够迅速收回成本。基于IOPS成本优势消除数据库分片架构 (1:16) 针对工程师普遍认为闪存价格昂贵的观点，Bergman辩称若以关键衡量指标——每IOPS的单价（gigabytes per IOPS）来评估，SSD实际上远比机械硬盘划算。他指出，关系型数据库的连接查询（joins）之所以缓慢，根本病因在于机械硬盘的随机I/O瓶颈，而非算法或数据库模型本身；业界为了规避缓慢的join而强推分片，属于因果倒置。单台现代服务器大约花费80,000美元即可配备50 TB的SSD容量，Bergman认为对于大多数数据量在50 TB以内的团队而言，彻底采用SSD足以直接免去复杂的分片方案。闪存存储在吞吐性能与能耗上的实际表现 (1:16) Bergman引用Wikia文件服务器的具体运行指标展示SSD的性能：该系统可在9分钟内遍历NFS上的800万个文件，12分钟完成全量备份，并稳定保持2.2 GB/s的纯随机写入吞吐量。从硬件单体性能对比来看，传统机械硬盘仅能提供约180 IOPS，而一块标准商用SSD即可达到约44,000 IOPS，单盘效能足以替代50块旋转磁盘。与此同时，固态介质的功耗从机械硬盘的每盘15瓦锐减至约1瓦，带来极高的能效收益。优先采用主流商用驱动器的渐进落地策略 (1:16) 在硬件选型策略上，Bergman建议企业避免盲目采购动辄耗资数万美元的小众SLC企业级加速卡，而应从主流大厂的基础商用级产品入手，例如单块售价约1,000美元的600 GB Intel 320系列。他形象地比喻道，对于目前还在“骑自行车”（机械硬盘）的团队而言，主流商用SSD就如同“法拉利”，完全没有必要越级追求难以承受且不实用的“F1赛车”。在实际管理中，SSD甚至已被Wikia和Fastly作为新员工入职时的签约礼品发放。 Bergman的核心论点在于彻底打破“闪存过于昂贵”的误区，强调通过SSD解决随机I/O短板可以从根本上重塑存储效率并简化系统架构。不过，演讲中的成本估算与免分片建议均建立在2011年当时的数据规模（50 TB以下）与硬件价格基准之上，未进一步展开探讨超大规模分布式系统的容灾备份及冷数据存储权衡。

Source quotes

[1:16] Artur Bergman: you can get 50 terabytes of ssds in one machine for like $80,000 and you can fucking skip sharding unless you have more than 50 terabytes [0:02] Artur Bergman: you're basically wasting your life like everyone who doesn't have an SSD in your laptop you're wasting your life

Original transcript

[0:02] hi um how many of you have ssds in your [0:08] laptops okay how many of you use ssds in [0:11] your data [0:13] centers kind of hard to see but okay how [0:16] many of [0:17] you are completely SSD in your data [0:20] centers except for backups and near [0:25] Long that's like one person and I built [0:29] that [0:31] so okay you're basically wasting your [0:34] life like everyone who doesn't have an [0:36] SSD in your laptop you're wasting your [0:37] life and this is the boot time of my [0:50] Mac right 12 seconds there are about [0:55] what 1,700 people here takes a minute to [0:57] boot your Mac every time we boot the Mac [1:00] every time you boot your computer we're [1:01] basically wasting a day of time right [1:06] so you know go buy an SSD put in your [1:09] laptop they're like 500 bucks and they [1:11] will save it back to you instantly and [1:12] if you run developer teams just go do it [1:16] so the problem is and the reason I'm [1:18] giving this talk is I've been saying [1:19] this for two and a half years and I keep [1:21] repeating that to every single [1:24] individual I talk to and what I get back [1:26] is they're too expensive actually [1:28] they're cheaper than drives [1:31] right because the relevant metric that [1:33] you need to look at is gigabyte per [1:37] iops right that is the important [1:40] part so wik just bought a new file [1:43] server 8 million files on an Fs in 9 [1:46] minutes full back 12 minutes and you [1:49] have 2.2 gabt per second of random [1:53] rights random rights like you can do [1:56] joints you can do whatever you want you [1:57] can get 50 terabytes of ssds in one [1:59] machine for like $80,000 and you can [2:01] [ __ ] skip sharding unless you have [2:03] more than 50 terab which is actually [2:05] more most of [2:07] you you know [2:09] like just not needed because the problem [2:13] with joints Is Random Access right I [2:17] hear what's the problem with joints [2:18] they're slow why are they slow because [2:20] they're joints why are they bad because [2:22] we can't CH why do you need to char [2:24] because otherwise it's slow no it's [2:25] because of [ __ ] this [2:27] guio right and it's one watt per c p CPU [2:32] instead of [2:34] 15 like that's where you can save power [2:38] it's the best way you can save power and [2:39] also you need like one SSD instead of 50 [2:42] hard drives actually one SSD is like [2:44] 44,000 iops and one hard drive is 180 [2:47] iops so do the math on how much you can [2:51] save start small just get like Intel [2:54] 320s don't get fancy SLC cards for like [2:57] tens of thousands of dollars you don't [2:59] need that you can't drive a [ __ ] [3:00] Formula 1 car and you're currently on a [3:03] bicycle right so just get a [3:06] Ferrari thousand th bucks 600 [3:13] gigs wrong way so thank you very much [3:19] um and [3:27] uh the uh the signing gift at Wikia to [3:31] anyone who joins is say uh SSD sorry [3:35] well at fastly probably at Wikia [3:37] too well it is [3:39] now all right thank you arur Kanye West [3:45] Bergman I'mma let you finish but ssds [3:48] are the coolest discs of all time