Shopify built its own telemetry platform, and its architect would now buy the managed version

The transferable part of this account is not the throughput figure but the schema argument. Shopify's team concluded that metrics, logs, traces, exceptions and profiles are one primitive, structured time-ordered high-dimensional events, and put all of them on a single engine so a single question could be asked across all of them ^[6]. Before that, asking one question across metrics, logs and traces meant stitching answers together by hand across separate vendors ^[12]. They began with metrics, where the pain was sharpest, then moved the rest onto the same engine ^[7]. That order is copyable at any size. The volume is not.

The arithmetic is worth doing, because the source leaves it undone. At 100 million events per second and roughly 110 GB/s uncompressed, the average event is about 1.1 KB ^[2]. Peak is twice steady state, so the headroom designed into the system is 2x, not an order of magnitude ^[1]. Hold peak for an hour and that is around 396 TB of uncompressed telemetry ^[3], set against the 90 PB Shopify says it stored across its fleet during Black Friday and Cyber Monday ^[9]. The query side is easier to place: engineers expect answers in under a minute ^[4], and a 16x gain implies the same lookup previously sat near a quarter of an hour ^[4].

What the post does not contain is money. Previously unpredictable vendor costs were brought under control ^[2], which is a direction rather than a figure; there is no prior bill, no current bill, and no headcount for the platform team that has been rebuilding this since 2021 ^[11][6]. The piece is published by ClickHouse, drawn from a talk at an event ClickHouse hosted ^[14], and it still carries engineering director Elijah McPherson's line that he would probably reach for ClickHouse Cloud if he were starting again ^[5].

Read that next to the original complaint. The 2021 problem was spend growing faster than the fleet and core production tooling tied to roadmaps Shopify did not control ^[11][12], and the engine won partly because being open source meant Shopify could contribute to it ^[13]. Self-hosting was the method, not the objective, and the person closest to it now prices the managed option differently.

The setting matters if you intend to model this against your own bill: about 500 Kubernetes clusters, 1.5 million pods across five continents ^[8], hundreds of producing teams and schemas that keep changing ^[15]. Somewhere in there, per-gigabyte vendor pricing loses to salaries. Shopify found its crossover. The published account gives you the volume side of that calculation in full and the cost side not at all.