Fink Alert Broker for Time-Domain Astronomy

• Fink Alert Broker is a modular system that ingests, annotates, and filters real-time astronomical alerts to create scientifically prioritized subsets.
• It employs a distributed architecture using Apache Spark, Kafka, and HBase to perform rapid quality cuts, cross-matching, and data enrichment.
• Its flexible, module-driven design supports diverse science applications, including early SN Ia, kilonova, AGN classification, and Solar System object characterization.

Fink is a community alert broker for time-domain astronomy, designed for the Vera C. Rubin Observatory Legacy Survey of Space and Time and operated in practice on the public Zwicky Transient Facility alert stream. In the literature, it is defined as a system that ingests alerts in real time, annotates them with catalogue cross-matches and machine-learning outputs, filters and redistributes science-specific substreams, and preserves raw and enriched products for later reprocessing and object-centric analysis (Möller et al., 2020).

1. Definition and historical role

Fink emerged from the broker problem created by Rubin/LSST-scale alerting. The foundational description presents it as a community broker whose task is to stand between the raw public alert stream and science users, converting a high-rate sequence of alert packets into tractable, scientifically prioritized subsets. Its scope includes automated ingestion, annotation, selection, redistribution, long-term storage, and reproducible reprocessing, with explicit attention to the fact that broker decisions affect follow-up selection functions and therefore later population studies (Möller et al., 2020).

The same literature places Fink in the official Rubin/LSST broker ecosystem. Benchmark work based on ELAsTiCC identifies Fink as one of the community brokers selected to receive the raw LSST alert stream and treats ELAsTiCC as the intermediate step between current ZTF operations and Rubin-scale deployment (Fraga et al., 2024). In operational terms, later module papers describe Fink as having processed the public ZTF alert stream since November 2019, using ZTF as both a live science platform and a Rubin precursor environment (Biswas et al., 2022).

Fink is also notable for its modular scientific breadth. Rather than implementing a single global classifier, it hosts science modules contributed by different groups, each optimized for a specific selection problem: early Type Ia supernovae, kilonova-like fast transients, active galactic nuclei, Solar System objects, anomaly detection, hostless extragalactic transients, superluminous supernovae, tidal disruption events, and orphan gamma-ray-burst afterglows. This makes Fink closer to a broker platform than to a monolithic alert-routing service.

2. Architecture and processing model

The original architectural description specifies a distributed system organized around four clusters: a processing cluster, a communication cluster, a science-portal cluster, and a data-store cluster. Real-time processing is performed with Apache Spark Structured Streaming; alert redistribution uses Apache Kafka; long-term storage uses HDFS; and object-centric portal access is backed by Apache HBase. Incoming ZTF packets are decoded, written to persistent storage, enriched by science modules, and then redistributed as filtered Kafka topics or exposed later through the science portal (Möller et al., 2020).

The same description makes clear that Fink was designed for near-real-time rather than per-packet serial execution. Its trigger policy combines a time-based and count-based rule and fires after 400 alerts have been received or 15 seconds have elapsed. For the ZTF public stream, Fink applies broker-level quality cuts before heavier science processing: RealBogus score above 0.55, number of prior-tagged bad pixels in a $5 \times 5$ stamp equal to 0, and absolute difference between aperture magnitude and PSF-fit magnitude below 0.1. These cuts remove roughly $70\%$ of the ZTF public stream, leaving about $30\%$ for downstream science modules (Möller et al., 2020).

Cross-matching is a core architectural service rather than an afterthought. The broker paper distinguishes three modes: large-catalogue positional cross-matches through the CDS xmatch service, in-house cross-match tools for smaller catalogues, and stream-to-stream association against live multi-messenger and multi-wavelength feeds received as VOEvents through Comet. The standard SIMBAD cross-match radius is 1 arcsecond. Persistent storage of both raw and enriched alerts is coupled to versioned software and replay tools, so that historical broker decisions can be reconstructed later (Möller et al., 2020).

ELAsTiCC-based LSST tests extended this architecture from ZTF-scale operations to LSST-like conditions. During the challenge, Fink decoded incoming packets, applied classifiers on every alert, and returned enriched packets to the DESC team. With 24 cores in parallel and nine deployed classifiers, $82.2\%$ of alerts were classified in $< 30$ s, $90.5\%$ in $< 1$ min, and $> 99.9\%$ in $< 10$ min. The same study reports that the challenge deployment used 33 cores total for listening, processing, and result return, and that the Fink Data Transfer service had streamed more than one billion alerts since challenge start (Fraga et al., 2024).

3. Science modules and broker-native classification

A recurrent theme in later work is that Fink’s scientific logic is implemented through broker-native science modules. In the NOMAI paper, such modules are described as community-built pipelines that consume incoming alerts, enrich them with derived quantities or classifications, and attach those products back to the alert stream (Russeil et al., 16 Apr 2026). This modular design supports both broad taxonomic classification and highly specialized rare-event sieves.

The early SN Ia module is one of the clearest examples of active-learning-driven broker logic. A Random Forest classifier trained through uncertainty sampling reached $89\%$ purity and $70\%$0 efficiency on real ZTF alerts, starting from an initial sample of 10 alerts and evolving through 300 iterations. In live broker operations from 01/November/2020 to 31/October/2021, Fink reported 809 early SN Ia candidates to TNS; among the 535 that later obtained spectroscopic classification, 459 were confirmed as SNe Ia (Leoni et al., 2021). A later real-time active-learning follow-up program used Fink to select the 10 alerts closest to $70\%$1 each night and obtained 92 spectroscopic classifications, yielding a training set that, with 25% less spectra, improved classification metrics when compared to publicly reported spectra (Möller et al., 26 Feb 2025).

The AGN module shows the same broker-oriented preference for lightweight but scalable features. Implemented within Fink, it uses summary statistics and symbolic-regression-based color estimation for sparse, irregular multiband data, together with active learning to build an efficient labeled training set. On real ZTF alerts, it achieved $70\%$2 accuracy, $70\%$3 precision and $70\%$4 recall (Russeil et al., 2022). The kilonova science module takes a different route, projecting light curves onto a principal-component basis and training a random forest to separate fast KN-like from slow non-KN-like events. Classification based on long light curves achieved $70\%$5 precision and $70\%$6 recall, while the broker-like 30-day regime achieved $70\%$7 precision and $70\%$8 recall (Biswas et al., 2022).

LSST-oriented classification benchmarks broaden this module portfolio further. On ELAsTiCC, Fink evaluated tree-based targeted classifiers and deep-learning broad classifiers, including CATS and SuperNNova. CATS achieved $70\%$9 precision for all classes except `long', for which it achieved $30\%$0, while the best periodic binary classifier achieved $30\%$1 precision and $30\%$2 completeness (Fraga et al., 2024). For rare luminous transients, NOMAI was deployed inside Fink as a real-time SLSN module using Rainbow and SALT2-derived features with XGBoost; on its archival benchmark it reached $30\%$3 completeness and $30\%$4 purity, and in its first two months of live evaluation it recovered 22 of the 24 active SLSNe reported on TNS (Russeil et al., 16 Apr 2026). A separate rising-phase TDE module, also based on Rainbow features and XGBoost, achieved $30\%$5 recall and showed that, among known TDEs passing the selection cuts, half were flagged before halfway in their rise (Lanza et al., 23 Jul 2025).

Fink has also been used to develop extremely selective rare-event filters. For orphan short-GRB afterglows in Rubin/LSST-like streams, a Scikit-Learn gradient boosting classifier was deployed in a Fink-like broker mode with decision threshold $30\%$6; on the held-out test set it retained 452 of 679 simulated orphans while only 1 of 10000 ELAsTiCC non-periodic events passed (Masson et al., 2024). For hostless extragalactic transients, ELEPHANT first operated as an archival Fink-based pipeline on ZTF, reducing $30\%$7 alerts to 1563 hostless candidates (Pessi et al., 2024). In later live Fink deployment between 1 September 2023 and 31 December 2025, ELEPHANT flagged 877 ZTF objects as hostless candidates and the paper reports an overall accuracy of 0.84 (Durgesh et al., 21 May 2026).

4. Solar System and object-level characterization

Although Fink is usually discussed in the context of transient brokering, the literature shows that it also functions as an object-level characterization platform for Solar System science. One branch concerns physical parameter inference from sparse multi-band photometry. The sHG1G2 phase-function model was implemented within Fink and run on ZTF alert-stream data to estimate $30\%$8, $30\%$9, $82.2\%$0 in each band together with spin-orientation and oblateness parameters. Using 13,245,908 observations of 122,675 Solar System objects, the study derived acceptable sHG1G2-based parameters for 95,593 objects, which it describes as about a tenfold increase in the number of characterized objects relative to the current census for spin-related properties (Carry et al., 2024).

The operational model for this SSO work differs from low-latency transient scoring. Fink ingests ZTF alerts daily, accumulates SSO photometry, computes SSO parameters once a month, and exposes the results through the Fink web portal, the Fink API, and the Solar System Objects Fink Table. This is a broker-mediated catalog-enrichment service rather than a per-alert veto or trigger, and it shows that Fink’s role includes persistent object-centric computation as well as real-time alert selection (Carry et al., 2024).

A second Solar System branch is discovery of previously unreported moving objects. Here the broker first reduces the full alert stream to a manageable Solar-System-candidate subset, and only then calls the linking algorithm Fink-FAT. Between November 2019 and December 2022, Fink processed 111,275,131 ZTF alerts and reduced them to 389,530 new Solar System alert candidates. Fink-FAT then extracted 327 new orbits from those candidates, of which 65 were still unreported in the MPC database as of March 2023 (Montagner et al., 2023). In this case Fink’s main contribution is not the orbit solver alone, but the broker-level preselection that makes real-time linkage computationally tractable.

5. Dissemination, human vetting, and follow-up workflows

Fink is designed not only to score alerts but also to expose them through operational interfaces suited to follow-up. Different papers describe access through the Fink client, bot-based substreams, Kafka live streams, the science portal, REST APIs, the Livestream service, the download interface, and dedicated Slack or Telegram channels. In the real-time active-learning follow-up study, the critical distinction was between streaming and archival access: the Science Portal and API were too delayed for prompt observing because the relevant database is populated at the end of the night, whereas the Slack-bot substream supported real-time candidate delivery. In that study, the median delay between observation receipt, ingestion, processing, and filtering was $82.2\%$1, with the 10th percentile at $82.2\%$2 and the 90th percentile at $82.2\%$3 (Möller et al., 26 Feb 2025).

The anomaly-detection module illustrates the most explicit human-in-the-loop workflow. Operating inside Fink on the ZTF stream, it transforms light curves into compact statistical features, ranks alerts with an Isolation Forest, and distributes the top 10 most anomalous candidates each night via Slack, Telegram, and the Fink REST API. A public Telegram bot collects expert feedback, and the ranking can be refined using Active Anomaly Discovery. In its first year of operations, beginning 25 January 2023, this broker-mediated workflow led to the discovery and follow-up of the rare AM CVn system Fink J062452.88+020818.3, the unusual transient with precursor SN 2023mtp, and the UX Ori-type star Fink J222324.32+744222.0, while also yielding 30 previously unreported supernovae and nine new dwarf novae (Pruzhinskaya et al., 31 Mar 2026).

A complementary example is the GRANDMA kilonova campaign, in which Fink served as the real-time decision engine for a networked follow-up program. Between 1 April and 30 September 2021, Fink received $82.2\%$4 ZTF alerts over 160 observing nights; after quality cuts, $82.2\%$5 alerts remained for science-module processing. Three kilonova filters then selected 107, 68, and 127 alert candidates, respectively. Six broker-selected candidates were followed by GRANDMA using 37 telescopes, and rapid follow-up constrained fading rates within $82.2\%$6 days post-discovery, without waiting for further ZTF observations (Aivazyan et al., 2022).

This dissemination layer is also used by newer modules. NOMAI posts nightly SLSN candidates to dedicated channels, and the TDE module publishes nightly candidate cards including cross-links, fit summaries, and classifier outputs. The operational pattern is consistent across these examples: Fink reduces alert volume aggressively, publishes compact candidate streams with broker-added context, and relies on expert review or telescope networks for the final transition from statistical candidate to astrophysical interpretation.

6. Rubin/LSST preparation, limitations, and outlook

ELAsTiCC is treated throughout the literature as Fink’s principal LSST rehearsal. The LSST benchmark paper concludes that Fink classifiers can handle the extra complexity expected from Rubin data, but it also states that transitioning from current infrastructures to Rubin will require significant adaptation of the currently available tools (Fraga et al., 2024). That dual conclusion recurs in module-specific studies: Fink is already capable of LSST-like streaming classification, yet many ZTF-era modules depend on assumptions about cadence, passbands, alert history length, or class balance that will need revision.

Some limitations are intrinsic to the science modules rather than the broker core. NOMAI requires at least 30 days of photometric history, more than 7 total observations, and more than 3 observations in each of $82.2\%$7 and $82.2\%$8, so it is explicitly not an ultra-early classifier (Russeil et al., 16 Apr 2026). The rising-phase TDE module requires at least seven points and rising/non-decaying behavior, which is scientifically appropriate but narrows the operational domain (Lanza et al., 23 Jul 2025). The orphan-GRB-afterglow filter is described as a first version of a classifier, restricted to $82.2\%$9, dependent on at least five light-curve points for feature extraction, and trained only against ELAsTiCC non-periodic backgrounds; even before machine learning, only about $< 30$0 of simulated orphan afterglows had at least one point in Rubin pseudo-observations (Masson et al., 2024).

Other limitations arise from current survey depth and contextual ambiguity. ELEPHANT’s ZTF-era hostless sample is affected by cataclysmic-variable contamination and by the fact that many apparently hostless events simply have hosts below ZTF depth or outside the analyzed stamp center, although the Rubin version has been adapted and has been processing Rubin alerts since February 2026 (Durgesh et al., 21 May 2026). In Solar System science, the sHG1G2 study notes that ZTF’s typical $< 30$1 observations per asteroid over about 3 years limit recovery of subtle phase and spin information, whereas Rubin is expected to provide roughly $< 30$2 observations per asteroid over 10 years (Carry et al., 2024).

The long-term trajectory nevertheless remains clear. Across the broker and module papers, Fink is consistently presented as more than a transport layer: it is an extensible, versioned, broker-native analysis environment in which real-time classification, historical reprocessing, object-level enrichment, and human-guided follow-up are all part of the same system. Its importance in the Rubin era lies less in any single classifier than in this combined capability to ingest, annotate, rank, redistribute, and retrospectively reinterpret an alert stream whose scientific content will be concentrated in a very small fraction of all nightly events.