Astronomer's 84 percent planning figure counts only the respondents still on Airflow 2

Two migration numbers in Astronomer's report do not share a base. 26 percent of respondents have finished the move to Airflow 3, which shipped in April 2025 ^[11]. The 84 percent is a share of the respondents still on 2.x, so the 58-point spread the dev.to write-up draws subtracts a subset's share from the whole sample's ^[12]. Put both on one base: 74 percent unmigrated times 0.84 gives about 62 percent of all respondents planning a move they have not started ^[13]. That leaves roughly 12 percent doing neither ^[13].

The breaking-change list explains the delay better than intent does. Airflow 3 removes SubDAGs, drops the execution_date, prev_ds and next_ds context variables, stops xcom_pull(key=) from searching upstream tasks, and cuts worker access to the metadata database so calls route through the REST API ^[14]. Three of the four delete a name a DAG referenced; the xcom_pull change alters what a call finds ^[14]. The write-up's worked example is 300 production DAGs with half using execution_date for partition logic, which it puts at weeks of refactoring before anyone touches CI ^[15]. Its four-person team at a Series B with 150 DAGs plans the work for Q3, pushes it to Q4, then Q1 ^[18].

The contributor counts are the figure being passed around: over 3,600 unique Airflow contributors, against Spark's 2,000-plus and Kafka's 1,530 ^[6]. The write-up converts that into 135 percent more than Kafka and 80 percent more than Spark ^[7]. Only Kafka's number is precise, and 3,600 over 1,530 is 2.35 ^[8]. Spark's is a floor set against a floor, so an actual Spark count of 3,600 would leave no gap ^[8]. The write-up credits Airflow's provider packages: cloud connectors, custom operators and executor integrations that lower friction for a first-time contributor ^[9]. It also calls contributor velocity the best leading indicator of whether a project keeps pace with the ecosystem around it ^[10]. Cumulative unique contributors counts everyone who has ever committed code, and the summary does not say over what period the three counts accumulated ^[6].

What the survey can measure is state inside Airflow's install base, and there it splits by who answered. 48 percent of Astronomer's own customers run Airflow 3, and 60 percent of those with more than 50,000 employees do ^[16]. The write-up attributes that to dedicated platform teams that can absorb migration work while the rest of the org keeps shipping ^[17]. Sample size does not repair the frame: 5,818 respondents ^[1] recruited around a managed Airflow platform cannot rank orchestration against compute, because the people running Spark or Kafka and no Airflow were never in the population.

Fieldwork closed on 20 November 2025 ^[2], more than four months before Airflow 2 reached end of life in April 2026 ^[24]. The 26 percent therefore describes the install base before the cutoff for security patches and provider package updates landed ^[19]. The sample itself grew 10.8 percent over the 2025 edition's 5,250 respondents, and added six countries ^[5].