By Alex Merced, Data Lakehouse and AI Evangelist
Release season arrived across the lakehouse projects this week. PyIceberg 0.12 shipped, Arrow Rust 59.3.0 passed its vote, Parquet Java 1.18.1 collected its binding +1s, Polaris lined up 1.8.0, DataFusion started backport triage for 55.1.0, and the Iceberg Terraform provider recovered from a failed release candidate on the strength of one careful license review. Underneath the release traffic sat a harder set of questions about who owns what. Iceberg contributors debated moving the DataFusion integration to a different PMC, Parquet contributors worked through what a format version number actually promises, Iceberg's catalog crowd argued over whether label metadata belongs to the catalog or the user, and the Ossie podling spent the week deciding what its own specification is for. Six projects, one recurring theme: boundaries.
Apache Iceberg
The headline release is PyIceberg 0.12, announced by release manager Alex Stephen on September 1 after the 0.12.0rc2 vote closed with binding +1s from Daniel Weeks, Kevin Liu, Fokko Driesprong, and Sung Yun, plus non-binding verification from Jared Yu, Xuanwo, Neelesh Salian, Yuya Ebihara, and Jean-Baptiste Onofré. The verification detail in that thread is worth reading if you have never watched an ASF release vote up close. Yuya Ebihara checked GPG signatures on every wheel and the source tarball on Apple Silicon with Python 3.13.3. Neelesh Salian ran the same checks on Python 3.12. Jean-Baptiste Onofré flagged a license issue in the source distribution, and Kevin Liu opened a GitHub issue to track it rather than block the release. Fokko Driesprong named conflict resolution as his favorite feature in the new version, which tells you where the Python implementation is heading. PyIceberg started as a metadata reader. It now handles write paths with the kind of concurrency semantics that used to require the Java library.
The Iceberg Terraform provider took a longer road. Sung Yun cast a binding -1 on RC2 after finding dependency drift and LICENSE-binary gaps that appeared between RC1 and RC2, specifically around the iceberg-go dependency. Jean-Baptiste Onofré added a 0 with a full checklist of what he verified in the source distribution. Matt Topol, the release manager, opened a fix PR the same day and cut RC3 two days later. That candidate collected +1s from Xuanwo, Alex Stephen, Fokko Driesprong, and Sung Yun, who confirmed that LICENSE-binary now covers the modules missed in RC2. This is the release process working exactly as designed. A binding voter caught a real licensing problem, the release manager fixed it in under 48 hours, and the first release of a new Iceberg subproject goes out clean.
The most consequential governance discussion of the week is the proposal to move the iceberg-datafusion integration out of the apache/iceberg-rust repository. Shawn Chang summarized the state of play after the community sync: the crate has a reviewer bandwidth problem that is slowing development. Gabriel Musat laid out the two facts driving the decision, namely the review bottleneck and the tight coupling between the crate and DataFusion's own release cadence. Renjie Liu pushed back on testing concerns as a blocker, noting that most of the sqllogictests live in PRs that modify the integration anyway. Then Xuanwo made the argument that seems to be carrying the room: DataFusion is the largest dependency of iceberg-datafusion, and Comet is its largest downstream consumer. Andy Grove added a +1 from the Comet side, calling the integration very important to that project. Renjie Liu closed the week by writing out the Comet and DataFusion relationship for people who do not follow the Rust side daily.
Read that thread alongside the Java release work and you see two different Iceberg ecosystems operating at different speeds. Neelesh Salian cleaned up the 1.12.0 milestone and reported on September 2 that the milestone is down to a handful of open PRs. Danny Jones from AWS volunteered as release manager for iceberg-rust 0.11 with Shawn Chang backing him on process, and Renjie Liu, Kevin Liu, Kurtis Wright, and Alexander Bailey all signed on within days. The Rust implementation moves fast and reorganizes itself when structure gets in the way. The Java implementation moves on a milestone and carries a decade of compatibility obligations.
Flink version support turned into the week's sharpest engineering debate. Péter Váry opened the question for 1.12.0 with a proposal to adopt an "LTS plus the two latest releases" support matrix. Maximilian Michels backed it as the approach that keeps Iceberg aligned with the Flink versions Flink itself supports at release time. Manu Zhang raised the production concern that matters to operators: Iceberg has never jumped two Flink versions at once, and teams running older Flink builds need an upgrade path that does not strand them. Talat Uyarer proposed a middle position that preserves version overlap, then went ahead and opened a PR to unblock the release candidate, with an explicit note that the PR does not preempt the discussion. Péter Váry asked Maximilian Michels how hard it will be to fold Talat's suggestion into the existing PR. The compromise looks close. The underlying tension is permanent, because Flink and Iceberg release on different clocks and every support matrix decision picks a winner between currency and stability.
On the spec side, the vote on finer grained read restrictions in the REST catalog opened August 31 and drew a wall of support: a binding +1 from Yufei Gu, and non-binding votes from Nevin Zheng, Andrei Tserakhau, huaxin gao, Alex Stephen, Gianluca Graziadei, and Holden Karau, among others. Prashant Singh drove this proposal through weeks of dedicated syncs, and the two loose ends closed just before the vote. The overlapping column projections question on nested types resolved toward disallowing overlap on both the catalog and client side. Prashant also added a compatibility kit for read restrictions so implementations have something to test against rather than a prose spec and good intentions. Column-level and row-level restrictions enforced at the catalog rather than the engine changes the security story for multi-engine lakehouses, because the restriction travels with the table instead of living in whichever query engine happens to be reading.
The catalog thread that generated the most heat was table and column label metadata in the REST catalog. Andrei Tserakhau brought the proposal back to the list after the catalog community sync, framing the motivating case as catalog-to-catalog federation. Prashant Singh raised concerns about labels used for governance decisions. Ryan Blue drew the line hard, rejecting the idea of reusing table config for this purpose and warning against mixing user-controlled properties into catalog-controlled configuration. That objection is the crux. Config already exists, labels look like config, and the difference is who owns the value and what trusts it. By September 2 Andrei and Prashant had narrowed the disagreement to how labels get surfaced in logging and AI context, with Andrei conceding that his earlier framing was too broad. Watch this one. Label metadata is the seam where catalogs stop being table registries and start being governance systems.
Several v4 metadata questions moved in parallel. Ryan Blue proposed a scheme for tracking field IDs for non-materialized columns, assigning table field IDs to values that never get written into the table. Gianluca Graziadei suggested splitting the treatment between expressions that are intrinsically deterministic and those that are not. Péter Váry raised schema evolution, pointing out that expression result types change as the schema underneath them changes, and noted the proposal makes index definitions much simpler. Sergei Nikolaev opened a separate v4 question on Avro timestamp types, arguing for dropping Iceberg-specific conventions in favor of the Avro 1.12.0 spec. Fokko Driesprong replied with the history, since he added the fixed[16] UUID encoding and timestamp-nanos to Avro in the first place. Andrei Tserakhau also split out a collations question about file prunability when engines run different ICU versions, which is the kind of problem that only shows up once you take multi-engine correctness seriously.
Smaller items worth your attention. Russell Spitzer explained the resolution of the EagerInputFile discussion: Varun moved the eager read into the Parquet file reader itself, so Parquet now buffers any file of one megabyte or less in a single request. For metadata-heavy workloads against object storage, request count is the cost driver, and collapsing small reads into one round trip pays off immediately. Rahul Mahadev proposed a standard User-Agent format for REST catalog clients so a catalog can tell which client is talking to it in a parseable way. Hongyue Zhang closed the loop on position deletes with row data, with the community agreeing to remove that handling from maintenance actions in 1.12.0. Tomohiro Tanaka responded to feedback on the table_properties_log metadata table, addressing read cost and server impact. Hemanth Boyina proposed extending validate-from-snapshot-id to MERGE so optimistic concurrency validation starts from a known snapshot the same way it does for overwrites. William Hyun and Sung Yun kept file-level access delegation in the REST spec moving, with Sung splitting Dan Weeks's two scenarios into three by separating presigned URLs returned from scan planning from those returned elsewhere.
The Iceberg community also picked up a new integration point from outside its own ecosystem. Gianluca Graziadei announced storm-iceberg, an external Apache Storm module that ingests streaming tuples straight into an Iceberg table from inside a Storm topology. Talat Uyarer scheduled a File Type sync for September 2 to align on the first-class file type proposal. And Danica Fine reminded both the Iceberg and Polaris lists that Lakehouse Day EU 2026 lands October 10 in Glasgow, co-located with Community Over Code.
Apache Polaris
Polaris spent the week getting 1.8.0 ready and settling three design questions that all trace back to the same root: how much of a table's physical layout does the catalog get to decide?
Jean-Baptiste Onofré volunteered to drive the 1.8.0 release at the beginning of September. Robert Stupp backed the schedule and argued for keeping the release focused on fixes and improvements that are ready today rather than stretching to include in-flight features. Jean-Baptiste agreed, noting that a monthly cadence means nothing needs to be forced into any single release. Yufei Gu made the case for one exception, PR 5053, on the grounds that the work is nearly done and has been in flight for a while. Dmitri Bourlatchkov seconded, and Jean-Baptiste agreed to take another review pass. Monthly releases change the psychology of scope negotiation. When the next train leaves in four weeks, nobody fights to get on this one.
Eundo Lee's work to make the relational JDBC schema name configurable is the community-health story of the week. Eundo sent a polite reminder that PR #4945 had been waiting while maintainers focused on the release. Alexandre Dutra responded within hours, proposed merging by end of day unless anyone objected, and then did. Yufei Gu registered a +0 with a real design reservation, arguing that the schema name should be a Polaris-owned property that Polaris maps to the backend rather than a passthrough, and chose not to block on it. Eundo folded in EJ Wang's review suggestions before the merge landed. A first-time-ish contributor got a clear answer, a dissenting reviewer said what he thought without stopping the work, and the feature ships in 1.8.0.
Yufei Gu opened the semantics of custom namespace locations by pointing out that ALLOW_NAMESPACE_CUSTOM_LOCATION currently skips parent-location validation entirely, so a namespace lands anywhere rather than under its parent. Dmitri Bourlatchkov challenged the premise that strict nesting applies universally, arguing that namespaces serve an organizational purpose that does not have to mirror storage layout. Jean-Baptiste went further and said the naming problem hides a deeper one, since whether tables follow namespace nesting is already governed by a different property, ALLOW_UNSTRUCTURED_TABLE_LOCATION. Two flags controlling overlapping behavior with names that suggest different scopes is a config bug waiting to become a security bug, and this thread is the right place to fix it before 1.9.0.
Alexandre Dutra also worked through forwarding user-defined principal properties in PolarisPrincipal with Prithvi S, reviewing the design doc, proposing an alternative, and signing off once Prithvi incorporated the changes. The threat model stayed constant through the discussion, with authorizers remaining trusted components. Alexandre returned to multiple StorageConfigurationInfos per catalog after a summer slowdown, picking up Srinivas's writeup. EJ Wang posted a status update on the Tag spec, with the API contract PR ready for review and a second PR covering definition CRUD. And Sung Yun reported that ASF Infra created the terraform-provider-polaris repository, scaffolded with LICENSE, NOTICE, a stub README, and an .asf.yaml.
Note the pattern across two projects. Iceberg is releasing a Terraform provider this week. Polaris just got its repository created. Infrastructure-as-code support for catalogs and table formats is becoming table stakes, because nobody wants to click through a UI to provision the governance layer of a production lakehouse.
Apache Arrow
Arrow's week centered on removing something rather than adding it. Antoine Pitrou pinged the dormant discussion about deprecating Tensor and SparseTensor messages in the IPC protocol, asking two questions: does anyone object, and does this need a formal vote? Curt Hagenlocher answered that a vote seems right, and made the observation that settles the substance. The only complete implementation of these messages is the C++ one, which means the feature does not meet today's bar for a format change. Matt Topol agreed on both the vote and the deprecation. Antoine opened the vote on September 1 and it filled up the same day with binding +1s from Curt Hagenlocher, Rok Mihevc, Micah Kornfield, Matt Topol, and Sutou Kouhei, plus non-binding support from Rusty Conover and Jacob Quinn.
Removing tensor support from a columnar format in 2026 reads odd until you remember what Arrow is for. Tensors in Arrow IPC never got multi-language implementations, so no cross-language interoperability existed to protect. Machine learning workloads that need tensor transport have their own well-supported paths. Carrying an under-implemented message type in the spec taxes every new implementation with no payoff. The word "informal" in the vote title matters too, since this deprecates without breaking anything that currently works.
Andrew Lamb ran the Arrow Rust 59.3.0 RC2 vote after RC1 hit a problem, and announced the result on September 1 with six +1 votes, four of them binding. Verification came in across Intel Mac from Ed Seidl, M4 Mac from Jeffrey Vo and L. C. Hsieh, and x86-64 Fedora 44 from Adam Reeve and Kosta Tarasov. The crate is published. The platform spread in that vote is quietly important, because arrow-rs sits underneath DataFusion, Comet, iceberg-rust, and a long tail of Rust data tools that all inherit whatever it does or does not verify.
Two format discussions stayed open. Mandukhai Alimaa posted a status update on the arrow.big_decimal extension proposal, noting that the effort started as a canonical arbitrary precision and scale type and has taken on more shape since. Micah Kornfield responded that a long pause is unnecessary and framed his own comments as pushing for conscious choices rather than rushed ones. Kosta Tarasov followed up on the Variant extension spec being inconsistent with the Parquet shredding spec and asked anyone working on Variant to take a look. That inconsistency is the sort of thing that becomes very expensive later, when two implementations both claim Variant support and disagree about what shredded data means on disk.
Wes McKinney revisited the status of Arrow's conbench data and the conbench open source project, asking whether anyone wants to engage before he takes a new codebase in a different direction under a different name. Rok Mihevc replied with a draft schema design and connected Wes with someone facing similar problems. Benchmark infrastructure rarely gets attention until it stops working, and Arrow's performance claims rest on the data it collects. Ian Cook also ran the biweekly community meeting on August 26.
Apache Parquet
Parquet is renegotiating its own compatibility contract, and this week showed the whole spread of that work: two votes passed, a patch release voted through, and one long-running design thread that hit a governance snag.
Alkis Evlogimenos closed the vote to remove self-references from the FILE logical type with seven binding +1s and eight non-binding, no dissent. Russell Spitzer, Antoine Pitrou, Micah Kornfield, Daniel Weeks, Ryan Blue, Julien Le Dem, and Gang Wu carried the binding side, with Rok Mihevc, Burak Yavuz, Divjot Arora, and Prateek Gaur among the non-binding voters in the vote thread. Divjot Arora followed with a second vote to specify handling for unrecognized logical and physical type combinations, where a reader drops to the physical type instead of failing. Andrew Lamb and Antoine Pitrou both voted +1 binding, with Antoine calling it short and useful. That is exactly the right instinct. Graceful degradation rules cost almost nothing to specify and save every future reader implementation from guessing.
Fokko Driesprong drove the patch release. After regressions surfaced in 1.18.0, he proposed a follow-up and cut 1.18.1 RC1 on August 31. Gábor Szádovszky verified the tarball, built and tested from source, checked artifacts with Steve's Auditor tool, and validated against Dremio builds and tests. Gang Wu covered signatures, KEYS, SHA512, tag and commit, a Maven Java 11 build, and license checks. Russell Spitzer, Peter Toth, and Gidon Gershinsky added their votes, and Fokko ran the release against an Iceberg PR to confirm nothing broke downstream. Testing a Parquet RC against Iceberg before it ships is a good habit, since Iceberg is where most Parquet regressions get discovered in production.
The versioning proposal is the thread to read in full. Ryan Blue summarized the open questions after a Wednesday call and followed up September 1 with a resolution on magic bytes, confirming with the modular footer authors that encryption can be handled within the footer, which frees the magic bytes to signal format version. Divjot Arora agreed with the summary and pushed back on using magic bytes for encryption signaling. Russell Spitzer endorsed the PAR3, PAR4 progression as the safest version signal. Micah Kornfield redirected the discussion toward requirements before logistics, arguing the community should agree on what the version number promises before deciding where the bits live.
Then Antoine Pitrou raised a process objection that deserves attention. He said the thread has become difficult to follow because it references discussions that happened privately or at least off-list, and that it is drifting from the original scope. He is right to say it, and the fact that he said it on the dev list is the system working. Parquet's design work now spans a versioning thread, a modular footer proposal, a Google doc from Julien Le Dem, a weekly sync, and a dedicated footer sync that Jiayi Wang ran on September 1 for its fifth session. Every one of those venues produces decisions. Only the mailing list produces a record. Projects that let the record fall behind the decisions lose the contributors who cannot attend meetings, which over time means losing the contributors who do not work at the two or three companies with the most people on the call.
On the encoding side, Andrew Lamb kept the ALP blog post review moving with Kosta Tarasov, adding BYTE_STREAM_SPLIT with ZSTD to the comparison after feedback from Antoine Pitrou, Arnav, Russell, and Jigao. Floating point compression in Parquet has been an open opportunity for years, and having measured numbers published in the open changes what implementations choose by default. Julien Le Dem ran the regular Parquet sync on August 26.
Apache DataFusion
DataFusion's week was short on threads and long on consequences. Tim Saucer opened the 55.1.0 release discussion, asking contributors to comment on the tracking issue with anything they want backported. Andrew Lamb crowdsourced the September ASF board report with a draft doc and a tracking ticket, which is the quarterly ritual that keeps a large, fast-moving project legible to the foundation.
The substantive thread is Comet. Andy Grove proposed a dedicated weekly sync call on the grounds that contributor count and velocity have both climbed sharply. Kazuyuki Tanimura, Kumar Ujjawal, Parth Chandra, and Marko Milenković all signed on. Bhargava Vadlamani proposed Friday mornings at 10:30 Pacific and offered to coordinate the meeting. Andy suggested documenting the call in the Comet contributors guide, and Bhargava agreed to open the doc PR and send the invite. A subproject earning its own weekly call is a growth signal, and Comet has been earning it. A Spark accelerator built on DataFusion sits at the intersection of the two ecosystems where most enterprise Spark workloads are heading.
Which brings the iceberg-datafusion question back around. The same thread about moving the integration ran on both dev lists, and the DataFusion side of the argument is straightforward. Comet depends on iceberg-datafusion, Comet lives under DataFusion, and the reviewers with the deepest context on both sit in the DataFusion community. Renjie Liu spent September 2 writing out the Comet and DataFusion relationship for readers who do not track the Rust ecosystem daily. If this move happens, DataFusion picks up ownership of a piece of the Iceberg stack, and Iceberg gains a maintainer pool it does not have to grow itself.
Apache Ossie
The Ossie podling is doing something none of the other five projects have to do anymore. It is deciding what it is.
Julian Hyde, who designed Apache Calcite and wrote the measures-in-SQL work that Ossie contributors already cite, introduced himself on the list after lurking for a while. Khushboo Bhatia welcomed him and explained that Ossie has been deliberate about keeping authoring concerns separate. Jean-Baptiste Onofré framed the project's purpose as standardizing the semantic model. Julian then asked the question that reframed the whole thread: if Ossie converts a semantic model from tool A to tool B, who decides whether the translated model means the same thing as the original? He offered one answer, which is to make Ossie a database-style API so semantics are defined by execution rather than by document structure. Matthew Mullins from Coginiti welcomed the pot-stirring, liked the proposal, and asked the community to do better than JDBC, which he described as inconsistently implemented everywhere. Khushboo replied that Ossie is not an interchange specification, that dropping the word "interchange" from the name was intentional, and that the repository now carries an expression language spec. Matthew's response was fair and pointed: if it is not an interchange spec, the README needs to stop reading like one. Will Pugh landed the synthesis, noting that an interchange without a definition of what it means cannot guarantee you get the same thing on the other side.
That exchange is the whole problem with semantic layer standardization in one thread. A semantic model is a set of definitions, and definitions only mean something relative to an evaluator. Two tools can accept the same YAML and produce different numbers for the same metric, and both of them will insist they implemented the spec. Ossie's answer to this is a compliance suite, which is why the PRs on foundational semantics and the compliance suite matter more than any feature PR in the repo. Khushboo argued for narrow, well-defined semantics that work end to end across many vendors over more powerful semantics that nobody defines precisely or adopts widely. Chris Eubank recapped the metric language working group meeting for people who missed it, and separately proposed seeding a BI SQL corpus as a follow-up action item from that meeting. A corpus of real BI SQL is how you find out whether a spec survives contact with what tools actually generate.
Release mechanics got their own debate. Jean-Baptiste opened the first Ossie releases discussion, and Yufei Gu argued for a source-only distribution since the community has not settled how to release individual converters. Yong Zheng asked whether the OSI to Ossie rename PR should land first, and Yufei agreed it should. Julian Hyde made the case for releasing the whole repository as one unit for the first few releases, noting that ASF releases carry legal weight and that the convenience of patching a single processor is a secondary concern. Yufei also kept the Python converter consolidation thread going, arguing that shared components exist even if release and review overhead cut the other way.
Vendor participation is picking up. Damian Waldron, a product manager at ThoughtSpot, brought the ThoughtSpot converter question to the list, covering scope, licensing, and where the code should live. Jean-Baptiste gave the standard ASF answer: Apache License 2.0, and code coming from another product needs a software grant agreement and a license change. Damian split the specification piece into a separate thread proposing THOUGHTSPOT be added to the Dialect enum, which is the right instinct, because adding a vendor to a spec enum is a different decision than accepting a code contribution.
Feature design ran hot on GitHub-backed discussions too. A proposal for shared filters, shared dimensions, and metric references drew eleven messages, with Josh Klahr digging into filter scope problems and cross-model references. Josh separately proposed dataset-scoped metrics, letting metrics be declared on an individual dataset instead of only on the semantic model, and Khushboo backed it while noting the subtleties belong on the PR. A spatial dimension type discussion picked up interest from CARTO, and a contributor from Databricks pushed on making relationship cardinality explicit and on how model-level semantic filters flow through consumers. Ankit Tandon posted notes from the September 1 Ontology working group sync.
The podling also grew. Jean-Baptiste announced Yong Zheng as a committer, and Khushboo Bhatia announced Josh Klahr joining the PPMC. Khushboo and Yufei both signed off on the September incubator report that Jean-Baptiste drafted.
Cross-Project Themes
Ownership beats architecture. The iceberg-datafusion move, the Parquet versioning debate, the Iceberg labels argument, and Ossie's interchange identity crisis are all the same class of question wearing different clothes. Nobody is arguing about whether the code works. They are arguing about who decides what it means and who has to maintain it. Xuanwo's case for moving the DataFusion integration rested on dependency direction and reviewer location, not on technical merit. Ryan Blue's objection to reusing config for labels rested on who controls the value, not on schema design. Ossie's whole week reduced to whether a spec that defines structure can promise meaning. As these projects mature, the interesting decisions stop being architectural and start being about boundaries between communities.
The verification bar keeps rising, and it is catching real problems. Sung Yun's binding -1 on the Terraform provider found license drift between two release candidates. Jean-Baptiste flagged a license issue in PyIceberg that Kevin Liu turned into a tracked issue. Gábor Szádovszky ran the Parquet RC through a third-party auditor tool and against downstream builds. Fokko tested Parquet 1.18.1 against an Iceberg PR before voting. This is not ceremony. Four of these projects ship artifacts that every other one depends on, and a bad release propagates through the stack in days.
Spec work is converging on compliance testing. Prashant Singh shipped a compatibility kit alongside the Iceberg read restrictions spec. Ossie is building a compliance suite as the mechanism that makes its semantics real. Divjot Arora's Parquet vote specifies reader behavior for combinations nobody anticipated. Kosta Tarasov flagged the Variant and Parquet shredding inconsistency precisely because two specs describing the same bytes differently is a bug that only testing surfaces. The community learned from a decade of table format ambiguity that a spec without a test suite is a suggestion.
Rust is where the reorganization happens. iceberg-rust is cutting 0.11 with a new release manager, splitting out its DataFusion integration, and shipping fast enough that structural changes get proposed and resolved inside a single week. arrow-rs shipped 59.3.0 with verification across four platform combinations. Comet is spinning up its own weekly sync on contributor growth. The Java implementations of these same projects operate on milestones, LTS matrices, and multi-year compatibility promises. Both models are correct for their constituencies, and the seam between them is where most of the interesting engineering now lives.
Semantic metadata is climbing into the catalog. Iceberg's labels proposal, Polaris's Tag spec, and Ossie's entire existence point the same direction. Teams want meaning attached to data at a layer that survives whichever engine reads it. The disagreements in the Iceberg labels thread are the early version of a fight every catalog will have, because once metadata drives governance decisions, the question of who writes it stops being a convenience question and becomes a security one.
Looking Ahead
Watch for the iceberg-rust 0.11 release candidate now that Danny Jones has consensus, and for a formal proposal on where iceberg-datafusion lands. The Iceberg 1.12.0 milestone is down to a few PRs, so an RC is close, with the Flink support matrix as the last real blocker. Polaris 1.8.0 should cut in the next week or two with PR 5053 included. Parquet 1.18.1 has its binding votes and needs only a result thread. Arrow's Tensor deprecation vote closes with more than enough support. DataFusion 55.1.0 is collecting backports, and the first dedicated Comet sync happens Friday morning Pacific. On the Ossie side, the first source release and the compliance suite PRs are the two things to track, since one makes the project real to users and the other makes the spec real to implementers.
If you want to go deeper on any of the topics in this issue, I keep a full catalog of my books on Apache Iceberg, lakehouse architecture, and agentic AI at books.alexmerced.com.
Top comments (0)