Latest / The Cloud Business Podcast with Fexingo: AWS, Azure, GCP, and Enterprise Infrastructure / Cloud Contract Now Mandates Write Amplification Factor Limits on SSDs
Transcript
- Lucas: So there's a clause creeping into cloud contracts that I think most teams haven't even noticed yet. It's about something called write amplification factor — WAF — and it's starting to show up in SSD storage provisions on AWS, Azure, and GCP. Luna: Write amplification factor — that's the ratio of physical writes the SSD actually does versus the logical writes the OS requests, right? Lucas: Exactly. Every time you write 1 megabyte logically, the SSD might internally write 2 or 3 megabytes because of garbage collection, read modify write cycles, and wear leveling. The industry average has been around 2x to 3x for modern drives. But now the hyperscalers are putting a hard cap in the contract — typically 3x WAF on their general-purpose SSD tiers — and if you exceed that, you get fined. Luna: Fined how? Is it a per gb penalty, or do they just throttle you? Lucas: It's a per gb surcharge on the excess writes. For example, on AWS gp3 volumes, the baseline is 3x WAF. If your application drives a 5x ratio, you're billed for the extra 2x as a premium per written gigabyte — sometimes up to 4 cents per gig above the threshold. That doesn't sound like much until you're doing petabyte-scale writes. Luna: Right — and that's not hypothetical. I recall a Google paper from late 2024 that found 40% of cloud SSD wear comes from noisy-neighbor workloads, not from the primary tenant. Lucas: That's the exact paper that kicked this off. The researchers showed that on shared storage nodes, one aggressive writer could accelerate wear on all adjacent volumes. So the providers said, 'We're not going to eat that cost anymore.' And now it's in the fine print. Luna: Who gets hit hardest by this? Like, what kind of workload drives high WAF? Lucas: Anything that does small random writes — think databases with frequent updates, logging systems that append in tiny increments, or container orchestration that constantly writes state files. Also, any application that does read modify write instead of full-page writes. A classic example: a Kafka cluster with default settings can easily push WAF to 4x or 5x because each small message write triggers a full 4KB page rewrite internally. Luna: So if you're running Kafka on gp3, you could be paying a 33% storage surcharge just from WAF penalties, on top of your normal compute. Lucas: Exactly. I talked to an ad-tech firm last month — mid-size, about 200 nodes — and they got hit with an $87,000 WAF penalty in a single month. Their engineering team had no idea the clause existed. They thought the contract was just standard egress and compute pricing. Luna: Ouch. So what's the practical fix? Can you tune your application to reduce WAF? Lucas: Yes — and this is where the clause actually drives better engineering. First, align your filesystem block size to the SSD page size. Most SSDs use 4KB pages, so if your application writes in 512-byte sectors, you're causing read modify write. Switch to 4KB writes. Second, use batching and buffering — accumulate small writes into larger ones. Third, move to log-structured merge trees for databases instead of B-trees, because LSM trees do sequential writes. Luna: So the clause is almost forcing people to adopt modern storage patterns they should have been using anyway. Lucas: That's the silver lining. But there's a darker implication I want to flag. Historically, cloud storage was marketed as 'you don't have to worry about hardware endurance.' Now, with WAF clauses, the provider is shifting that risk back to the customer. If your storage wears out faster because of high WAF, you pay for it — not them. Luna: Wait — does that mean the provider is essentially saying, 'We'll provision the SSD, but if you kill it faster than expected, you owe us'? Lucas: Exactly. It's a quiet reversal of the traditional cloud value prop. I've seen a few analysts call it 'the end of the no-worry storage era.' And the clauses are spreading fast. AWS introduced it in gp3 in 2025, Azure followed with Premium SSD v2 in early 2026, and GCP added it to Persistent Disk just last quarter. Luna: So every hyperscaler now has some form of this. What about provisioned IOPS classes — do those have different thresholds? Lucas: Yes — higher-end tiers sometimes have more generous WAF limits. For example, AWS io2 Block Express allows up to 5x WAF before penalties kick in, because those volumes are designed for high-write workloads. But you're paying a premium for that tier anyway. The real trap is on the standard tiers where most teams run their production databases. Luna: So when you're negotiating a contract renewal, what should you ask for? Lucas: First, ask for the WAF baseline in writing — it's often buried in the service terms, not the main contract. Second, negotiate a grace period: maybe 90 days where you can measure your WAF before penalties apply, so you have time to tune. Third, ask for a tool — some providers now offer WAF monitoring dashboards, but they're opt-in. If your account manager doesn't mention it, you won't get it. Luna: And if you're an SRE or a cloud architect, the first step is just measuring your current WAF. Most teams don't have that metric today. Lucas: Right. And you can measure it using the provider's own metrics. On AWS, it's the VolumeWriteAmplification metric in CloudWatch — but it's not enabled by default. You have to explicitly turn it on. I bet 90% of AWS customers don't even know it exists. Luna: Let's talk about that ad-tech firm again — how did they fix their WAF issue after the $87,000 penalty? Lucas: They did three things. First, they moved their Kafka broker logs from gp3 to a cold tier that doesn't have WAF penalties — basically a cost trade-off. Second, they increased the Kafka batch size from 512 bytes to 4KB, which dropped their WAF from 4.7x to 2.1x. Third, they switched their Cassandra SSTable compaction strategy from size-tiered to leveled, which reduced write amplification significantly. Within two months, their WAF penalties went to zero. Luna: So the fix was non-trivial but not insane — it took real engineering effort, but it was doable. Lucas: Exactly. And now they have a competitive advantage because their storage costs are lower than competitors who haven't tuned yet. This is one of those clauses where the early movers win. Luna: It reminds me of the data egress clause wave a few years ago — everyone ignored it until they got a surprise bill. WAF feels like the next egress. Lucas: It really does. And just like egress, there's a whole ecosystem of third-party tools emerging to monitor and optimize WAF. I've seen startups like SSDGuard and WriteWise launching specifically for this. The cloud cost optimization space keeps getting more granular. Luna: One thing I wonder about — on the provider side, are they actually enforcing these clauses aggressively, or is it still more of a scare tactic? Lucas: From what I've seen, they enforce it. The ad-tech firm I mentioned got a bill, not a warning. And I've heard from a few FinOps consultants that AWS is running automated WAF audits quarterly and back-billing for the past 12 months. So it's real. Luna: That's aggressive. So if you've been running a high waf workload for a year without knowing, you could be facing a six-figure retroactive bill. Lucas: Exactly. That's why I think every cloud engineer should check their WAF metrics this week. Not next quarter. Luna: We talk a lot on this show about the hidden complexities of cloud infrastructure — and this is a perfect example. If these conversations are useful for what you're building or running, that's exactly why we keep them ad-free and accessible. Lucas: Yeah, listener support is what makes that possible. If you find value in episodes like this one, you can support the show over at buy me a coffee dot com slash fexingo. Every bit helps us stay independent and dig into the details that matter. Luna: And we really appreciate it. Now — back to WAF. There's one more angle I want to touch on: what about on-premises versus cloud? If you own your SSDs, you don't have a WAF clause, but you do have to replace drives sooner. Lucas: True. And the cost of replacing an enterprise SSD is around $200 per terabyte, plus downtime. So the cloud clause is essentially forcing you to internalize that cost as a variable expense instead of a capital expense. It changes the TCO equation significantly for high-write workloads. Luna: So the decision to stay on-prem or move to cloud now has to factor in WAF penalties. That's a new variable that didn't exist two years ago. Lucas: Exactly. And I think we'll see more of this — cloud contracts becoming more like utility tariffs with surcharges for 'bad behavior' from the provider's perspective. WAF is just the beginning. Luna: Well, next episode we're planning to look at the new GPU memory bandwidth guarantee clauses — that one's also heating up. Thanks for listening, and we'll catch you next time. Lucas: Take care, everyone.