Summary
- The sentence that defines the Confluent Unit for Kafka calls each unit "a fixed amount of ingress, egress, partitions, and client connections"; the section beneath it classes ingress, egress and total client connections as recommended guidelines, and partitions and storage as fixed limits. Three of the four things the definition calls fixed are, in the operating section, advice. What is fixed is the number of units and the hours.
- My division of the page's own cluster maxima gives exactly 152 in five of six rows (9,120 ÷ 60, 27,360 ÷ 180, 2,736,000 ÷ 18,000, 76,000 ÷ 500, 2,280,000 ÷ 15,000), and 152 is the ceiling the same page names for Google Cloud while naming 252 for AWS. The sixth row, partitions, divides to about 22.2 — against a linear-scaling sentence that would allow 4,500 × 252 = 1,134,000 partitions, which is 11.34 times the 100,000 cluster maximum printed a few lines above it.
- No Confluent-authored page among the eight documents cited here prints a rule for combining per-dimension requirements into a unit count, and the public price list prints hourly rates for the elastic unit by tier while routing the fixed unit to "Contact us". A buyer can compute the units and the hours; the rate that turns them into money arrives by correspondence.
The sentence, and the section under it
Confluent Incorporated, whose directory entry records the company behind the product, defines the unit in three sentences on its cluster-types page: "A Confluent Unit for Kafka (CKU) is the unit of throughput capacity for Dedicated clusters in Confluent Cloud. Each CKU provides a fixed amount of ingress, egress, partitions, and client connections. You scale a Dedicated cluster by adding or removing CKUs."
Further down the same page, in a section headed "Fixed limits and recommended guidelines for Dedicated clusters", the four nouns are sorted into two classes. Ingress, egress, total client connections and requests are recommended guidelines. Partitions, storage, connection attempts, Kafka REST Produce v3 and Kafka REST Admin v3 are fixed limits. Only one of the four dimensions the definition calls fixed — partitions — is a fixed limit in the operating section. The other three are things you are told you can exceed.
The page's own account of what a guideline is: "The recommended guideline for a dimension is calculated for a workload optimized across the dimensions, enabling high levels of CKU utilization as measured by the cluster load metric. You can exceed the recommended guideline for a dimension, and achieve higher performance for that dimension, usually only if your use of other dimensions is less than the recommended guideline or fixed limit."
That is the closest thing to a sizing rule on the page, and it does not tell you how many units to buy. It tells you the numbers are joint: the ingress figure assumes you are not also sitting on the egress figure. The adverb carries the load — "usually only if".
Where "fixed" actually lands is the billing dimensions page: "A CKU is a fixed unit of pre-allocated capacity for Dedicated Kafka clusters in Confluent Cloud. Each CKU bundles throughput, partition, and connection limits. You set the number of CKUs at cluster creation and pay per CKU per hour." Pre-allocated, set at creation, billed per hour. The quantity fixed by the contract of sale is the count and the clock.
What one unit prints
The eCKU/CKU comparison table on the same page gives the Dedicated column: 60 MBps ingress, 180 MBps egress, 4,500 partitions before replication, 4,500 compactable partitions, 18,000 total client connections, 500 connection attempts per second, 15,000 requests per second, and 50 MBps maximum throughput for Kafka REST Produce v3.
The Enterprise eCKU column beside it prints the same 60 and 180, the same 18,000 connections and the same 500 attempts — then 3,000 partitions instead of 4,500, 1,000 compactable instead of 4,500, and 7,500 requests instead of 15,000. The throughput rows do not separate the two units at all. The partition and request rows do.
Six rows, five denominators
The page also prints cluster-level maxima: 9,120 MBps ingress, 27,360 MBps egress, 100,000 partitions before replication, 2,736,000 total client connections, 76,000 connection attempts per second, 2,280,000 requests per second. It prints CKU ceilings as well: 252 on AWS, 152 on Google Cloud, 100 on Azure.
Divide the maxima by the per-unit column. The arithmetic is mine: 9,120 ÷ 60 = 152. 27,360 ÷ 180 = 152. 2,736,000 ÷ 18,000 = 152. 76,000 ÷ 500 = 152. 2,280,000 ÷ 15,000 = 152. Five rows, one number. The sixth gives 100,000 ÷ 4,500 = 22.2.
152 is not an anonymous figure. It is the ceiling the page names for Google Cloud, printed a few lines away from 252 for AWS. No cited page derives the 9,120 row, or explains why the cluster maximum should equal the per-unit column multiplied by one provider's ceiling rather than by the largest ceiling the page prints. I am not alleging that anything was copied. I am saying that a reader who multiplies arrives at one cloud's number, and that no document tells him why that is the number.
The partition row fails in the other direction. The same page says: "For Dedicated clusters, these per-CKU limits scale linearly as you add or remove CKUs. For example, each extra CKU increases the Kafka REST Produce v3 connection limit by 300 requests per second." Scale linearly, then: 4,500 × 252 = 1,134,000 partitions, which is 11.34 times the 100,000 the page prints as a cluster maximum. At the Google Cloud ceiling of 152 the same multiplication gives 684,000. The linear rule and the cluster maximum cannot both bind. The page prints both and reconciles neither.
The formula that only a tutorial prints
To buy the right number of units you need an aggregation rule: do the dimensions add, or does the largest one win? Nobody at Confluent prints one. Six of the eight documents cited here are Confluent's own — the cluster-types page, the billing dimensions page, the billing FAQ, the billing overview, the price list and a Confluent Developer episode — and none states how per-dimension requirements combine into a unit count.
The only document I located that prints the rule outright is a third-party capacity-planning tutorial, which gives "Required CKUs = MAX(Ingress CKUs, Egress CKUs, Partition CKUs, Connection CKUs, Other Capacity Constraints)" and then rounds up for headroom. It is not a Confluent publication. The vendor's nearest approach is a peer-reviewed VLDB paper by Confluent authors, which describes the unit as exposing maximum capacity on each dimension rather than a summable budget — a description in a paper, not a sizing instruction in a product document, and one I can only cite as a description because retrieval returned a snippet rather than the page.
The spoken version is older and blunter. A 2022 Confluent Developer episode has staff saying, "In general, you are not able to max out all of the dimensions at the same time." The same episode cites 50 MBps ingress per unit, not the 60 MBps in the table above: the sentence outlived the table it was written against, and its number did not.
Two units priced in public, one not
The price list prints hourly dollars by tier for the elastic unit: Basic free for the first unit then $0.14, Standard $0.75, Enterprise $1.75 to $2.25, Freight $2.25 with a two-unit minimum. It prints per-gigabyte ingress and egress rates and per-gigabyte-month storage. For the unit this article is about, it prints a sentence: "We also offer Dedicated clusters available by pre-provisioned capacity for use cases with distinct networking and throughput requirements. Contact us to learn more or try for free."
The billing FAQ draws the line the price list avoids: the CKU charges for provisioned capacity, the eCKU for actual usage, and a Dedicated resize takes the new value from the following hour. The billing dimensions page adds that the elastic charge is the maximum used inside each hour, so two units then three means three. The billing overview introduces a third object entirely: "A Confluent Consumption Unit (CCU) is a billing unit used in marketplace listings, calculated from data ingress, egress, and storage."
Three units, two measurement bases. The CKU and the eCKU count capacity held or used; the CCU counts data moved. A customer comparing a marketplace listing to a Dedicated invoice is comparing a flow to a stock, with no printed rate to convert one into the other. The cost line item for the first two is KAFKA_NUM_CKUS, defined as cost per CKU/eCKU per hour — a numerator whose value, for the Dedicated case, is printed nowhere.
What a reader can and cannot compute
A reader can compute the unit count and the hours exactly. The relationship is linear, the count is a setting, and the hourly charge is the count multiplied by a rate. What no cited page supplies is the rate, or the rule for choosing the count in the first place.
Eight addresses are cited here. Two more were requested and returned nothing usable — a Dedicated-cluster URL that served the documentation landing page, and a quotas page that returned only its introduction — so neither is cited for any claim. The pricing page was retrieved in two renderings, both printing elastic rates and routing the fixed unit to sales. No contract, service-level agreement or order form was among the documents obtained, so the record shows what the fixed-limit class is called and what it is applied to, and stops there.
The units are arithmetic: exact, linear, printed. The bill is correspondence. Only one of the two can be finished without asking Confluent for a number.
Member Briefing
Read the full trend briefing
Members read the full briefing with the source notes behind every finding.
Strategic Circle
A daily briefing in your language, built from changes we have verified.
Join for USD 20/monthCancel anytime.
Members also get the full research library, a daily briefing on the networks they follow and the private forum.






