How we scaled metrics aggregation by moving durable state into storage
Comments on “How we scaled metrics aggregation by moving durable state into storage”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Datadog engineer Khayyam Guliyev describes three generations of its metrics aggregation service, each built for roughly 10× more volume. Static partitions with local RocksDB state doubled compute and needed manual splits. Dynamic Kafka assignment with blob-backed state made recovery the bottleneck. The third generation makes streaming best-effort and has storage reconcile partial aggregates, duplicates and late data, so workers resume at the live edge. The figures are Datadog's own, unsourced, and the evidence for the recovery benefit is one production incident.
Why it ranked
No. 19 of 142 with a score of 0.64. The panel rated it highest on new information (0.93) and technical depth (0.85), lowest on audience breadth (0.49). Placed No. 19, No. 46 by score alone: the placement terms moved other stories past it.
- AI judgeJudge 2claude-haiku-5-5
- Detailed engineering write-up on metrics aggregation architecture evolution with concrete recovery and state-management tradeoffs; useful to data and infrastructure builders, though company-specific and not a landmark.}
- AI judgeJudge 3gpt-5.6-luna
- A detailed production engineering account of scaling stateful metrics aggregation, covering partitioning, RocksDB recovery, autoscaling, deduplication, late data, storage-side reconciliation, and operational trade-offs across three generations.
Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.
Score
0.64total under rubric v2.3
- Significance0.57
- Reader interest0.76
- Audience breadth0.49
- Practical relevance0.71
- Substantiated0.65
- Technical depth0.85
- New information0.93
- Counts against: Promotional0.10
- Counts against: Overclaim0.13
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| SignificanceHow much does this change what people in tech know or can do? | 0.58 | 0.40 | 0.72 | 0.57 | 0.32 | +0.24 | +0.136 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.69 | 0.70 | 0.90 | 0.76 | 0.21 | +0.26 | +0.198 |
| Audience breadthHow many people who build technology does it concern? | 0.52 | 0.40 | 0.55 | 0.49 | 0.15 | +0.13 | +0.064 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.71 | 0.60 | 0.82 | 0.71 | 0.22 | +0.22 | +0.156 |
| SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations. | 0.49 | 0.60 | 0.86 | 0.65 | 0.37 | +0.05 | +0.032 |
| Technical depthHow technically substantive is it? | 0.86 | 0.80 | 0.88 | 0.85 | 0.08 | +0.05 | +0.042 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.91 | 0.90 | 0.97 | 0.93 | 0.07 | +0.05 | +0.046 |
| PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional. | 0.06 | 0.20 | 0.04 | 0.10 | 0.16 | −0.19 | −0.019 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.14 | 0.20 | 0.06 | 0.13 | 0.14 | −0.11 | −0.015 |
| Sum of contributions | +0.642 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.642 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-11-18/19/"><img src="https://rubric.hrih.si/badge/2026-11-18/19.svg" alt="Ranked #19 on Rubric, 18 Nov 2026" width="245" height="28"></a>Claim check
Claim check: 1 matches source · 6 no source given · 1 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
No source givenNumber
Each new generation of Datadog's metrics aggregation system had to support roughly 10× the metrics volume of the generation before it.
The article gives no link or citation for it.
The article says “Each new generation of Datadog’s metrics aggregation system had to support roughly 10× the metrics volume of the generation before it”
The source says No source to compare with.
Couldn’t checkCausal conclusion
Removing unnecessary tag dimensions reduces the volume of indexed custom metrics, which can lower costs and improve query performance while preserving mathematically accurate results across remaining dimensions.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “Removing unnecessary tag dimensions reduces the volume of indexed custom metrics, which can lower costs and improve query performance while preserving mathematically accurate results across the dimensions that remain”
The source says No passage could be compared.
Source cited by the article: docs.datadoghq.com/metrics/metrics-without-limits
Matches sourceGeneralization
DDSketches can be merged without storing every individual measurement while preserving accurate percentile calculations.
The cited source says the same thing.
The article says “These sketches can be merged without storing every individual measurement while preserving accurate percentile calculations”
The source says “Furthermore, by always using the same representative values, we make the sketch mergeable; merging is done by simply summing the respective counters of the sketches”
Source cited by the article: datadoghq.com/blog/engineering/computing-accurate-percentiles-with-ddsketch
No source givenCausal conclusion
In the first generation, the leader-follower model effectively doubled aggregation compute because both nodes performed the same aggregation work.
The article gives no link or citation for it.
The article says “This effectively doubled aggregation compute and made horizontal scaling difficult”
The source says No source to compare with.
No source givenNumber
One documented manual partition split in the first generation required approximately 1.5 hours of coordinated engineering work.
The article gives no link or citation for it.
The article says “One documented split required approximately 1.5 hours of coordinated engineering work”
The source says No source to compare with.
No source givenNumber
After migrating to the second generation, Datadog could support approximately 10× higher metrics volumes with minimal operational intervention.
The article gives no link or citation for it.
The article says “After the migration, we could support approximately 10× higher metrics volumes with minimal operational intervention”
The source says No source to compare with.
No source givenFirst, only or fastest
The third generation was Datadog's first use of its Streaming Platform for an aggregation service.
The article gives no link or citation for it.
The article says “This generation also marked our first use of Datadog’s Streaming Platform for an aggregation service”
The source says No source to compare with.
No source givenCausal conclusion
During a production incident, an operator reset processing to the live edge via the Streaming Platform UI while storage handled deferred aggregation work, showing that restoring current traffic no longer required reconstructing historical state.
The article gives no link or citation for it.
The article says “Restoring current traffic no longer required reconstructing all historical aggregation state first”
The source says No source to compare with.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN —
Recorded for context only, checked 18 Nov 2026, 05:47 UTC. It played no part in the ranking.