Changelog

What changed for the people who use AnnZarro and run its servers, newest release first. Each release says what you can now do, what behaves differently, and what to change when you upgrade. The AnnZarro paper describes v0.4.1; this log starts with the release after it. Older releases are listed on GitHub.

Unreleased

0.7.1 (2026-10-19)

Added

  • Plot Options, Axes and ticks: Flip y and Flip x run an axis the other way, for example so y grows downwards as it does in the image of spatial data. The flip is per plot, stays through zoom, Reset axes and double click, works with Equal aspect, the scale bar, large-plot mode and exports, and is kept in sessions, links and panel sets. Apply to other panels, presets and the session default leave it alone. 2D plots only.

Changed

  • Tables no longer print a note under the Advanced Search box listing every column with more than 10,000 distinct values. Equals and Not on such a column still take a typed value instead of a list.

0.7.0 (2026-10-19)

Added

  • Plot Options, Apply to other panels: Make this the session default makes a plot’s look what Default means for the session, per plot kind (cell or gene). Plots on Default (made from the chooser, or put back with the Default preset) take it now and follow it when you change or reset it (Reset to built-in default); plots with a look of their own, and plots saved before this, keep theirs. A plot saved on Default opens on Default in whatever session it opens in. The default travels with the session: a reload of the tab, the autosave, share links and saved panel sets carry it, and opening a link or a panel set with its panels open makes it yours (Add panels closed does not); Close all with the cache-clearing tick ends it.

  • Panel sets: a set with a default style says so on its card (Style). A set with no panels can be saved when a session default is set, and is only a style (Style only, one Apply style button, no dataset): a way to share a house style.

  • Line selection in the table panel: the Line filter button before a table’s search box lets you draw a line on a plot of cells (a cell table) or of genes (a gene table). The side to the right of the line (from the mouse press to the release) is filtered out: the table gets the plot’s axis columns, a derived column y − m·x and one condition on it, joined to the filter by an outer AND. The line and the shaded side are previewed while you drag, Shift snaps to 45°, Esc stops, 3D and large plots are left alone, and a 2D plot with a text axis shows a small tooltip saying why. While the tool is on, the table has a red frame and a badge in the dataset bar (click it, or press Esc, to stop); closing the table, Close all, loading a panel set or switching dataset also stop it. A derived column leaves the table when no condition uses it any more. Several lines make an area. New tutorial: Cut away part of a plot with a line.

Changed

  • A plot with a table filter set no longer refuses to draw when it goes into large-plot mode: it is drawn whole and the filter is ignored, and the Large plot tag says so. The plot keeps its table setting, so the filter applies again when a subset takes it out of large-plot mode.

  • Plot Options, Presets: Default now resets the whole look that Apply to other panels copies, including Not in table, Focused cell and the export size. It used to leave those three as they were.

Fixed

  • A cell table that showed “not available in large-plot mode” and was then given a subset to list (all cells, then a subset, then all cells, then a subset again) kept the note until you pressed Apply Changes: its previous DataTable, detached from the page by the note, received the new rows. The table now lets that table go when the note replaces it, so it lists the subset’s cells by itself.

  • In a table’s search row, the help of Regex, Smart and Case Sensitive shows just above its own button; it was drawn above the Advanced Search box, out of sight. The row’s buttons no longer touch the panel’s edges.

  • A plot coloured by a number and linked to a table drew the wrong points when the table’s filter matched nothing: the cells that had just left the table vanished instead of turning grey, and only a reload corrected it. Every cell now turns grey, and removing the condition brings the colours back.

0.6.2 (2026-10-19)

Added

  • Container recipes for hosting a lab server, in packaging/container/: a Dockerfile (multi-stage, non-root, gunicorn, health check, remote extra by build argument), a Compose example with nginx and TLS, and an annzarro.def for Apptainer on a cluster. The new page Deployment, Containers gives the commands. A CI job builds the image on pull requests that touch these files.

  • annzarro serve starts the production server in one command: gunicorn with the bundled configuration, reading the same site file and ANNZARRO_* variables as annzarro start, with --config, --host, --port, --workers, --threads and extra gunicorn options after --. It prints the URL, workers, threads, whether login is on and the configuration file, and login stays on unless the configuration turns it off. The new server extra installs gunicorn: pip install 'annzarro[server]'. The explicit gunicorn command still works.

  • An optional Prometheus /metrics endpoint (server.metrics_enabled, off by default): requests and latency by route group, requests in flight, response bytes, dataset cache hits, misses and size, open datasets, memory and the worker’s process id, merged across gunicorn workers. Only addresses in server.metrics_allow (loopback unless you list others) can read it; everyone else gets the reply of a URL that does not exist.

  • server.log_format: json (or ANNZARRO_SERVER_LOG_FORMAT=json) writes log lines on standard error as JSON, with a request id (also sent as X-Request-Id), method, route group, status and duration on each request’s access line. The default text format is unchanged. New page: Deployment, Monitoring a server.

  • GET /api/v1/health answers {"status": "ok"} without a login, for a service manager’s, container’s or load balancer’s probe; /api/v1/status still asks for one.

  • With JSON logs or metrics on, an admin change’s audit record carries the request_id of its request, the id on its log lines; the Audit log section shows it as the time’s tooltip. Older records verify as before.

  • Plot Options, Presets: Bare draws only the points (no margins, legend, colour bar, scale bar, axes, grid or zero lines; in 3D no axes or background planes), for figures and overlays. Default undoes it.

Changed

  • A numeric column of 8 MB or more is sent from its array in 1 MiB pieces instead of being copied into one bytes object first, so the server holds less while it sends: at 95.6 million cells the server’s peak memory in the first plot is 4.4 GB (v0.6.1: 4.8 GB). annzarro start also hands any large reply to its server in pieces. The bytes sent are the same.

  • A plot with the colour None draws its points blue (a lighter blue on the dark theme) in the regular mode too, as large-plot mode already did. Grey now means only “not in the table” (or a colour that could not be read, which the status line says).

  • The legend title sits above the entries at every legend position, starts where the labels start and has a clear gap below it; before it was flush left, a few pixels off the symbols, 2 px above the first entry (and beside the entries in a legend on top or at the bottom).

  • Admin page, Server health: each worker’s memory, and the total. Every worker now reports every 2 seconds whether or not it is serving, so a quiet worker is no longer shown as idle, and one that has not reported for a minute (not ten) is shown as stale: it is stuck. Server health and /metrics share one snapshot file and one writer per worker.

  • annzarro start serves through waitress (a new dependency, pure Python, also on Windows) instead of Flask’s development server. Debug mode and HTTPS still use the development server. A request line and headers over 262,144 bytes are refused with 431. It serves 32 request threads; server.threads (ANNZARRO_SERVER_THREADS) sets the number.

  • The lab-server guide’s nginx block keeps its connections to gunicorn open (upstream with keepalive, an empty Connection header) and speaks HTTP/2 to browsers. A plot loads through many small requests: a 4-panel view opened 184 connections to gunicorn per load on nginx before 1.29.7 and none with the block. With HTTP/2 and 150 ms of added latency the view drew in 4.1 s instead of 7.7 s. Its keepalive_timeout is set below gunicorn’s 5 s so nginx never sends into a closed connection.

  • The guide and the hosting checklist no longer tell you to start gunicorn with --limit-request-line 8190: it overrode the bundled configuration’s unlimited request line and refused the long URLs the web client builds. nginx’s large_client_header_buffers bounds the line.

Fixed

  • The search box in Load Panel Set had no right edge and a darker left edge, which showed most in Safari. The box now has one even border, and its focus ring follows the same outline.

  • A stream of every cell or gene name built the JSON of a whole stored chunk at once, which held about 35 MB more per reader than the chunk itself needs (a store with 2-million-name chunks). It is built 65,536 names at a time now; the bytes sent are unchanged.

  • Changing the subset while an earlier change is still loading shows the new choice at once: the header names the cells asked for, and only the last request’s cells are drawn.

  • The tooltip of Plot Options, Presets, Publication no longer says it sets the scale bar: it leaves the scale bar as it is.

  • Close all writes the autosave and the address at once, so a reload right after it no longer finds the panels just closed.

  • Switching Show Legend off or on in Plot Options redrew the graph once per legend entry: about a second on a 1M-point plot with 35 categories. It is now a single layout change (about 50 ms).

  • Every request a plot made opened a new TCP connection that the server closed after the reply, so a few servers loading plots side by side used up the machine’s ephemeral ports (“Can’t assign requested address” for other programs). Connections are kept open now.

Docs

  • The lab server page and the hosting checklist no longer tell you to pass --limit-request-line 8190 to gunicorn: it capped the request line that the bundled configuration leaves unlimited, so a plot that names many genes in its URL (about 12 KB for 2,000) is refused with 400 under it.

  • The README and the licence page cite v0.6.1 by its DOI, 10.5281/zenodo.23461950.

0.6.1 (2026-10-18)

Legend columns, dot size and title for categorical colours; the focused cell drawn as an open circle by default; a Background group in Plot Options in place of the colour theme buttons; a double-click that zooms without moving the focus; and fixes to the Plot Options menu, its opacity boxes, the site footer and the Structure search.

Added

  • Plot Options, Legend and colour bar: the legend of a categorical colour can have its title switched off, its dots set to a size of their own (2 to 16 px, independent of the points’ size) and be laid out in 2 to 8 columns. Unset, the legend looks as before. The options travel with links, panel sets and Apply to other panels.

Changed

  • The focused cell or gene is drawn as an open circle five times the point size by default (no fill, a 2 px edge, opaque). Switch No fill off in the Focused group of Plot Options to fill it; plots saved before this draw the new default unless they set the fill.

  • The Appearance group of Plot Options is now Background (background colour and 3D backdrop). The Plot colours buttons (Follow app theme, Light colours, Dark colours) are gone: plot colours follow the app theme until you set one, and each colour’s Reset hands it back. Saved links, panel sets and autosaves that pinned a palette open and draw as saved.

  • Double-clicking a plot to reset its zoom no longer changes the focused cell or gene. A single click now focuses after Plotly’s double-click window (0.3 s).

  • The site’s footer line on the main page is at the bottom of the page, under the panels, instead of pinned to the bottom of the window: scroll down to reach it.

Fixed

  • The arrow keys of the opacity boxes in Plot Options (Focused, Not in table) step from the opaque default (1, 0.95, 0.9) instead of jumping to the minimum; the boxes show the opacity in use (the Not in table one shows the automatic value, and stores nothing until you change it).

  • The Plot Options menu no longer closes when you press inside it (to select text in a box, drag a slider or a colour) and let go outside it. It closes on a click that starts outside.

  • When the server does not send the libraries the page needs (an installation missing files in static/vendor), AnnZarro says which ones and what to do, instead of “Can’t find variable: bootstrap”.

  • The Structure tab of Dataset information: the tree no longer shows through the strip above the search box while scrolling.

Docs

0.6.0 (2026-10-17)

An About dialog and what a public site needs (imprint, privacy notice, a footer), set from the admin page; a Dataset information dialog with attribution, README and structure; last sessions listed instead of redrawn, and plots of a set too large for the browser refused one by one instead of crashing the tab; plot options that survive links and Duplicate; and a Publication preset for figures.

Added

  • Last sessions. The welcome screen and the Load Panel Set dialog list up to three last sessions first (“Last session (autosaved 14:32)”, the dataset and the panel count), each with Restore. A session whose tab ran out of memory while it was drawn is marked “did not finish loading” and offers Restore with the default subset; it is never restored by itself again.

  • No fill in the Focused cell group (Plot Options) draws the focused marker as an open circle, its edge only, in 2D, 3D, gene and large plots, in exports and in links, panel sets and Apply to other panels. An edge width under 1 px is raised to 1 px when it is switched on, so the circle always shows.

  • The Structure tab of the Dataset information dialog has a search box: it filters column names, keys, uns paths, types and category labels as you type, marks the matches, opens the sections that hold them and counts them. A .* switch reads the query as a regular expression, Aa matches case (both remembered); an invalid or too slow pattern is said in words and the last result stays. Escape clears it, / and Ctrl/Cmd+F focus it.

  • uns["annzarro"] is a versioned format (schema annzarro.dataset-info, schema_version 1.0; a dictionary without them is version 1). The reader ignores keys it does not know, reads a newer minor version in full and a newer major version for the fields that did not change (the dialog says so), and every field accepts a simple form and a richer one: a licence as {id, url, text}, authors with ORCID iDs, several contacts, funding and provenance. annzarro store annotate writes the marker, takes --from-json for the richer forms, and refuses a value the reader would ignore. The field table and the compatibility rules are in the deployment guide.

  • A Dataset information button beside the open dataset’s name shows the dataset’s attribution (title, licence, authors, citation with a copy button, DOI, source, contact), its README rendered as safe Markdown (raw HTML is shown as text, links only http, https and mailto, no images) and its AnnData structure (cells and genes, obs and var columns with types, flags, category labels and colours with a warning when the colour count differs from the category count, obsm, layers and the other slots with shapes and, for sparse matrices, density, raw, unknown groups, the uns tree). Each tab loads when first opened, with a loading state and a retry on failure; the dataset browser’s details pane shows the same for a dataset that records it. Dataset owners write it into uns["annzarro"] and uns["README"], with annzarro store annotate for a zarr store that is already written; CELLxGENE’s uns["title"] and uns["citation"] are shown too. See “Dataset information and attribution” in the deployment guide.

  • Click the app name at the top left to see which version you are running, how to cite AnnZarro, and who develops and funds it (the Setty Lab at Fred Hutch Cancer Center, with the NIH award numbers), with links to the documentation, the repository and the changelog.

  • A server can show who runs it and what it owes its visitors. Under site: in the configuration, set an imprint, privacy notice, accessibility statement and terms of use, each as a link, a Markdown text or a Markdown file, plus an operator, a contact and the link labels (Impressum, Datenschutz for German sites). A small footer line on every page, including the sign-in, account and admin pages, links them; the About dialog gets a “This site” block, and a click on the imprint opens it in place. The texts are also served at /site/imprint, /site/privacy, /site/accessibility and /site/terms, which need no sign-in. Nothing appears until you configure it.

  • site.robots tells search engines to stay away (by default on a server with login), site.security_contact serves /.well-known/security.txt, and site.storage.persistent_convenience: false keeps the layout autosave and recent datasets only until the tab closes.

  • The Gene Set panel’s question before it sends gene lists now says that each service also receives your IP address, and links the service’s privacy policy.

  • site.software_credits: false reduces the About dialog’s credit for AnnZarro (lab logo, funding, citation) to its name, version, licence and links; it stays by default everywhere, below the site’s own block.

  • The About dialog of the desktop app states what the app sends and where.

  • The server warns at start when integrations.external_requests is "on" on a shared server, and when branding.contact_info.custom_html contains a script, an iframe or an address on another server.

  • Public site on the admin page: every site: setting (imprint, privacy notice, accessibility statement, terms, operator, contact, labels, footer, search engines, security contact, language, the AnnZarro credit, browser storage) can be set while the server runs, and the footer, About and the /site/ pages show it at once. Each field explains itself on hover and links the public-site guide, a Markdown text has a live preview, a value the server would refuse at start is refused under its field, and a checklist says which of imprint, privacy notice, contact and accessibility statement are missing. Each text shows what is in effect and where it comes from, a text the configuration file reads from a file with its path and as the pages show it. A save sends only what you changed: looking at another option, or a link or text left empty, changes nothing, and Not shown over the configuration file’s text says that it hides it. An admin’s value wins over the configuration file until Reset to config file; policy.locked: [site.imprint] keeps a key in the file’s hands. A text’s file can only be set in the configuration file.

  • Data & privacy (/site/data), generated from the running configuration: what the browser keeps, the cookies, what the Gene Set panel may send to which services and when, and that there are no analytics. Without a configured privacy notice it says plainly that it is not one. /site/about shows the About dialog as a page.

Changed

  • Every setting on the admin pages has an info mark and a Guide link. Hover or tab to the mark for what the setting does, its default and what it affects; Guide opens the explanation in the docs. This covers the Settings tab and the per-user settings, the Public site, the fields of the Users and Access panels, and Data sources. The admin page’s introduction now states the real rule between the config file and the admin: an admin value overrides the file unless the key is in policy.locked, and a restrict key can only be narrowed.

  • A new visit no longer redraws the last session by itself; a reload of the tab still does. A panel set kept on “All cells” and moved to a very large dataset could run the tab out of memory, and every reload drew it again. ui.autosave.auto_restore now takes reload (the default), always (the old behaviour, also true) or never (also false); ui.autosave.show_in_list lists the last sessions and is on by default. Upgrade: set auto_restore: always to keep the old behaviour.

  • The address names the open dataset. Every way of opening one (the dataset browser, the back button beside its name, a panel set, Restore, the desktop app, the dataset opened at start) writes ?dataset_path= (its path in the data directory, such as ?dataset_path=bm_aging.zarr), with no panel open too, so a reload opens the dataset on screen. The address is replaced, not added to the history.

  • The bare address opens no dataset. A fresh visit to it no longer opens the dataset opened last; the welcome screen lists the last sessions and the recent datasets, with Browse datasets…. A data directory with only one dataset still opens it, and the desktop app still reopens the dataset opened last.

  • The subset size you choose is kept for the tab. It still follows you from dataset to dataset and through a reload of the tab, but a fresh visit (a new tab, the address typed, a bookmark) starts on the server’s default subset. Upgrade: a size kept by 0.5.5 is removed at the first load, not carried; choose it again once in each tab.

  • Plots drawn at once are drawn in turn (a restored session, a dataset switch or a larger subset with plots open): each is checked against what the plots before it hold, the browser’s allocation check included, so the later ones of a set too large for the browser are not drawn instead of running the tab out of memory. Each says what ran short (“Not drawn: not enough browser buffers”, “… JS memory”, “… browser memory”, or “too many WebGL canvases”) with Half the cells, Subset…, Close other plots and Draw anyway (with what it risks). A Cell or Gene Table of more rows than fit is not loaded, with the same choices.

  • The Publication preset sets the font to Arial and draws a shorter colour bar with fewer ticks (40 % long, three ticks); Default puts the font back.

  • The panel’s text font now reaches the legend (title and items) and the colour bar (title and tick labels) as well as the axes, when you change it in Plot Options, press a preset, and in exports.

  • Plot images are exported at twice the size set in Plot Options by default (2400 × 1600 pixels for the default 1200 × 800), so a figure stays sharp in print; the modebar camera follows the panel’s export size. Untick Export at 2x for files of exactly the pixels set. A panel or link saved with the scale switched off keeps it off; one that never recorded it exports at 2x. SVG files keep the size set.

  • A plot’s toolbar (zoom, pan, reset, camera) sits in the top right corner of the plotting area instead of the figure’s, so a legend or colour bar placed on the right, the default, no longer runs into it.

  • A colour bar placed on top of or below a plot lies wholly outside the axes, with the margin it needs made for it, instead of over the top or bottom of the plot. It sits below the x-axis title at the bottom. A legend above or below a plot does the same, clear of the axes and the x-axis title, in short panels too.

  • The About dialog says that the Setty Lab develops and funds AnnZarro (the software), so that a site that runs it is not mistaken for the lab’s own.

  • A server other people reach (login on, a network address, server.hosted or a proxy) shows a footer line even with nothing configured: “About · Data & privacy”, plus the configured texts and contact. The desktop app and a personal server show none until you configure something.

  • Such a server logs a warning at start when it publishes neither an imprint nor a privacy notice, naming the admin page’s Public site section and the guide.

Fixes

  • A dataset README longer than the 200,000 characters the dialog renders says that it was cut (“Showing the first 200,000 characters of … KB”); its end used to be left out without a note.

  • The dataset browser’s path field opens with / or ~ (the first character of the path, as in the macOS and GTK file dialogs), from the list or from an empty search box; a search box that holds text keeps the character. The hint reads “Type a path or URL: /” on every platform. It advertised Cmd+L, which Safari keeps for its address bar and never passes to the page; Ctrl+L and Cmd+L still work where the web browser passes them on. In the list, / no longer moves to the search box.

  • Choosing “All cells” on a dataset the server shows in full anyway (up to its subsampling threshold) no longer carries every cell to the next, larger dataset: the kept size is forgotten and large datasets open on the server’s default subset. Every cell chosen on a larger dataset is still carried.

  • The subset dialog counts the array buffers of every open plot together. With six Cell Plots open on a 95.6-million-cell dataset, a computer with 128 GB let every cell through: its memory holds six large plots of 8 to 9 GB, but one browser tab holds about 16.6 GB of array buffers in Chrome and about 15 GB in Safari, and the tab ran out. Such a size is now refused with “Needs ~60 GB of browser buffers”, and the dialog says up to which size the plots fit (about 26 million cells for six plots in Chrome, 23.6 million in Safari). A new or redrawn large plot is checked the same way beside the plots already open. ui.memory.buffers_gb sets the limit; by default it is the browser’s own.

  • After the tab ended while drawing large plots, the next page keeps the buffers it held as this browser’s limit, so the same plots at the same size are not offered as fitting again. Before, a large-plot draw never left the crash marker that makes the guard more careful.

  • The margin learned from a crash no longer applies to everything, forever. It grew by 0.25 with every crash, for every action of any size, and was never forgotten. It now applies only to actions of at least half the size of the one that crashed, is forgotten after a week or once an action that large succeeds, and a refusal that only it causes says so and offers Reset the margin (in the panel and the subset dialog). A margin kept by an earlier version is dropped.

  • A new Cell Plot was refused although the browser had the memory, and allowed when it did not: the allocation check’s answer was inverted (since 0.4.2) for a new or duplicated large plot, a full-resolution export, a recolour and hover labels of more than 256 MB. After closing a plot of every cell of a 95.6-million-cell dataset, or on a fresh page that opened it on every cell, the create tile said “Not enough browser memory” with nothing open, and kept saying it after the subset was reduced to 100,000 cells.

  • The create and duplicate tiles are checked again when the number of cells shown changes (a smaller subset, another dataset), and a few seconds after a plot is closed, once its memory is freed. A tile refused at every cell stayed refused at 100,000 cells until another panel was opened or closed; a click on a refused tile also checks again.

  • Every memory refusal names the limit that binds: the create tile’s note, the panel’s headline and its status tag say “JS memory”, “browser buffers”, “browser memory” or “WebGL canvases” instead of “browser memory” for all of them.

  • The dataset information dialog shows the dataset’s name on one line, cutting the start with an ellipsis (the full path is its tooltip), and tables and code in a README use the dialog’s full width.

  • A dataset’s README, and the texts of a public site, render bare web addresses as links, and pipe tables, fenced and indented code blocks, horizontal rules, underlined headings and block quotes instead of showing their markup. Tables scroll sideways inside their own box when they are wider than the dialog, and show at most 500 rows and 30 columns.

  • A category hidden with a click on the legend stays hidden when a Cell Plot linked to a filtered Cell Table removes or greys out the cells not in the table. Hiding moved to another category, so other cells disappeared and came back, and a category the table filter emptied came back shown. The hidden entries now follow their category, or colour group, through every redraw, in regular and large plots, until you colour by another column.

  • A plot opened from a share link, a panel set, the autosave or Duplicate now looks as it did where it was made. Its axis lines, which the first draw took from the Grid switch, were missing whenever Grid was off, which is what the Publication preset does; so were the colours of the legend title and the toolbar, the tick colour, and Show Legend (the legend and the colour bar came back). Changing an axis, text or grid colour, or the font, in Plot Options now redraws every element that uses it, as opening the same plot does. Links made by earlier versions need no repair: the settings they carry were right, only the first draw ignored them.

  • A new focus shows sooner in big Cell Plots. At 1M cells a click showed the cell’s name after about 0.2 s and its marker after about 0.4 s, and kept the page busy for about 0.2 s; now both show within about 60 ms. Moving the marker no longer redraws every point through Plotly, and an open Cell Table no longer lists all its filtered rows each time the focus changes or the layout is autosaved; the periodic autosave also waits for a quiet moment, and leaving the page still saves at once.

  • A categorical colour whose uns["<column>_colors"] has fewer colours than the column has categories no longer repeats them, which gave different categories the same colour. As in scanpy, the stored colours are then set aside and the default palette is used, in regular and large plots, and the panel’s status strip says so (“Default colours: see why”).

  • Clicking a cell in a Cell Plot focuses it every time. In plots of hundreds of thousands of cells, a click soon after the plot was drawn or resized, or while the previous focus change was still being shown, often did nothing. In large plots, a click made while the plot is still preparing click selection (which can take a while at tens of millions of cells) shows a busy cursor and “Preparing click selection…” in the status strip, and focuses the cell once it is ready, instead of doing nothing.

  • A Cell Table or plot too large for any browser page (95.6 million rows against about 4 GB of JavaScript memory) now says “Too many rows for this browser” with the numbers, and offers “Show N cells” for the largest subset that fits (a table at most 1,000,000) and “Subset…”. It no longer offers “Close other plots” when no other plot is open or closing them cannot help, nor “Draw anyway” above the browser’s ceiling.

Docs

  • New page Running a public site: who is responsible, what a public site owes its visitors (imprint, privacy notice, accessibility), what AnnZarro stores and sends, the site settings, and privacy-notice templates in English and German.

  • The README and the licence page cite v0.5.5 by its DOI, 10.5281/zenodo.23440020.

0.5.5 (2026-10-17)

Fewer pop-ups, with messages shown where you are looking; automatic point size and opacity that follow the panel’s size; a subset size that follows you from dataset to dataset; and Plot Options for the focused cell’s marker.

Added

  • Plot Options has a Focused cell group (Focused gene in gene plots) that sets the size (a multiple of the point size), opacity, fill colour, edge colour and edge width of the marker on the focused cell, in 2D, 3D and large plots. The colours follow the app theme until you set them (the edge is black on a light plot and white on a dark one); the values are kept in panel sets and share links, and a link without them draws the usual red dot.

Changed

  • Automatic point size and opacity follow the plot area as well as the number of points: a large plot (a full window, a big export) draws larger, more opaque points and a small split pane smaller, fainter ones, so points look about as dense at any panel size. They update when the window is resized, a split is dragged, the controls are shown or hidden, or a panel is maximised; a size or opacity you set stays as set.

  • The cell subset size you choose follows you to other datasets. A dataset you switch to opens on that many cells when it has more, and with every cell when it has as many or fewer; its badge then reads, for example, “All cells (≤ 10,000)”, and the next larger dataset is subsampled to the size again. The size is kept per browser and user, so a reload keeps it, and the dataset AnnZarro opens when it starts uses it too. “Close all” with the tick box for clearing everything forgets it. Share links, panel sets and the restored session still open on the subset they saved.

  • A plot or table whose data the browser ran out of memory for says so in plain words (“the browser ran out of memory – Close a plot, or show fewer cells (a smaller subset), and try again”) instead of “failed to read – Array buffer allocation failed”, and offers Half the cells and Subset… where it says it.

  • Fewer pop-ups: AnnZarro now pops a notice up only for what you did not cause or cannot see (an error, data that changed under you, a partial outcome). Messages about what you just did appear where you are looking. A panel set that cannot be saved, loaded, imported or deleted says why on its card or under the name box in the dialog, and saving over an existing name turns Save into Overwrite and deleting asks on the card, in place of the browser’s own confirmation windows (also for Clear closed panels). A plot export or copy that the memory check refuses, or that fails, says so in the plot’s status line, with the advice, until the next one succeeds. What the Log colour scale does to a Min or Max you typed is shown under the boxes. Share Link is off, with a tooltip, until a dataset is open. No notice any more for a restored autosave, an imported panel set, a plot source missing from the dataset (the menu and the plot already say it), or a dataset that became unavailable (the dataset browser says it). Several columns removed from tables in one restore, or two notes about one store, come as one notice, and notices that stay up are held while the pointer or keyboard focus is on them and last longer when they are longer.

Fixes

  • The Save Panel Set dialog now closes when a save finishes while it is still fading in (a quick save, or Overwrite pressed at once); it used to stay open.

  • A plot coloured by nothing, or by a column the dataset does not have (after switching datasets), no longer shows a grey “trace 0” in a legend after an axis change; such a plot has no legend.

  • An error notice is no longer wiped by the notices that follow it from the same action (a shared link that could not be read, a cell subset that could not be applied): a notice is replaced only by one at least as severe, and a lower one waits below it.

  • The first notice after the page loads is announced by screen readers, and a dataset that fails to open is announced once, not twice.

  • A release build no longer fails when Apple’s ticket service drops the connection while the notarization ticket of the macOS app or its dmg is being checked (stapler validate); the check is retried up to five times, and a missing ticket still fails at once.

Docs

0.5.4 (2026-10-16)

A filtered Cell Table opens about twice as fast on large datasets, and switching datasets no longer shows a notice: a back button beside the dataset name returns to the previous one. The first release the desktop app can install over 0.5.3 by itself.

Changed

  • Switching datasets no longer shows a notice. A back button beside the dataset name returns to the previous dataset, and then offers the one you left, so the two toggle. The notice’s button used to spread across the page when a dataset name was long.

Fixes

  • A release build no longer fails when Apple’s ticket service drops the connection while the macOS app or its dmg is being stapled: stapling is retried up to five times.

  • A Cell Table with a filter opens about twice as fast on large datasets, from a share link, a panel set or a reload: the filter goes over the rows once instead of twice, and each row costs less. Landing on a link with a filtered table of 484,125 cells held the page for two stretches of over 6 s each (long enough for the browser to report the page unresponsive); the longest is now about 2 s.

Docs

0.5.3 (2026-10-16)

The desktop apps of 0.5.2 with a fix for Windows. 0.5.2 was published on PyPI only; its desktop apps were not released, because the Windows build found this bug. Everything in 0.5.2 below is in 0.5.3.

Fixes

  • Windows: the dataset browser’s details, the folder listing and the recent datasets work for a store on another drive than the data directory (for example the data directory on C: and the store on D:); the details answered with a server error.

0.5.2 (2026-10-16)

A dataset browser replaces the dataset dropdown, AnnZarro has a dark theme that follows your system, and the desktop app updates itself, installs with Homebrew, opens files from the File menu, Finder and drag and drop, and uses the AnnZarro header as its title bar. Admins control remote datasets in the admin interface.

Dataset browser

  • The dataset is a button at the left of the bar under the header, in place of the dropdown, and Ctrl+O (Cmd+O) opens the dataset browser from anywhere, also in a text field. The header keeps only the focus boxes and the panel set buttons, so a long dataset name does not squeeze them. The button names the open dataset (its path relative to the data directory, cut in the middle to fit, the whole path as its tooltip) and shows when a dataset is opening, when a view has no data (“No data: name”, Choose…) and when the open dataset became unreadable (“Not available”). The bar shows the dataset’s format beside its cells and genes, and names it once.

  • The browser lists Recent, Pinned, the data directory (and the directories an admin added, or a restricted user’s grants) and, when the server allows remote datasets, Remote, which says what can be opened and takes a URL. The path above the table turns into a path or URL field with Ctrl+L. The table is sortable (names ignore case and read numbers by value: liver_2 before liver_10), filters by format, number of cells and date, tags the datasets that match a view you are choosing for, and a Show more button pages it. Clicking a name opens the dataset; clicking elsewhere in the row selects it. The search (substring or regular expression) stays inside the folder shown.

  • A details pane (a sheet below 1100 px) shows a dataset’s counts, format, size, columns, obsm keys and layers, and runs the store check for a local store within its time budget; a remote store has a Check chunking button. Open, Pin and Copy path sit at its top and do not move while the details load. When panels are open, it names the keys they use that the selected dataset lacks.

  • A click on the dataset button, or Ctrl+O, while the page is still starting is kept: the button shows it is busy and the dataset browser opens once the page is ready.

  • After a switch a message, “Opened name”, has a Back to old name button. Opening a dataset needs no confirmation.

  • Opening with nothing asked for: the dataset you opened last, if the server still lets you read it, else the only dataset there is, else none, with the welcome screen offering Open a dataset (your recent ones and Browse datasets…). A server with several datasets does not open the first one alphabetically any more, which on a shared server could be a very large store.

  • Recently opened keeps stores in subfolders and remote stores (s3://, https://), shows one you can no longer open greyed in place, and is kept per server and per user: two people who sign in to one server from one browser do not see each other’s list. The list kept before moves to the first user who signs in afterwards, once.

  • The browser tab and the window are titled with the open dataset, bm_aging_annzarro.zarr · AnnZarro (with your app_name), and with the app name alone when none is open.

  • New optional setting branding.admin_contact (a name or an email address): a user who sees only some datasets is told “Ask contact for access.” when a link or panel set names a dataset they cannot read.

  • Keyboard and screen readers: the table is a native table with a caption and sortable header buttons, one tab stop with arrow keys, Home, End, type-ahead, / for the search field and Escape back to the dataset button; counts and openings are announced; the focus ring is a 2 px outline in the link colour in both themes; forced colours, reduced motion and touch targets (44 px) are respected; below 768 px the dialog fills the screen.

  • Select2 (the dropdown’s library) is removed from the page and the licence notices.

Dark theme

  • AnnZarro follows your system’s light or dark setting, live: switch the system while the page is open and the app switches. The System / Light / Dark buttons at the right end of the dataset bar override it for this browser and are remembered there. The login page, the admin pages, the account page and the desktop app are themed too, and there is no flash of the wrong theme on load.

  • Plots follow the theme for every colour you have not set (background, grid, axes, text, legend, colour bar, the toolbar, the grey Not in table points and the scale bar). A colour you set, or a plot theme button, is kept when the theme changes. Plot Options > Appearance has Follow app theme (and Light colours / Dark colours); a colour’s Reset follows the theme again. The Default preset follows the theme, Publication stays white.

  • Share links and panel sets store only the colours you set. A link from an earlier version opens with its colours equal to the old light defaults following the theme; the colours you chose, and the old Dark Theme button’s, stay.

  • Exports use the colours on screen: export in the Light theme (or with Publication) for a paper.

  • Grid and axis colours apply from the first draw, and several grey-on-white text and control colours are slightly darker for WCAG AA contrast.

Desktop app

  • Updates itself. It looks for a newer release on launch and then once a day, asks before it downloads (Download, Later, Skip this version) and again before it restarts, and does nothing without your say-so. Check for Updates… is in the AnnZarro menu (macOS) or Help menu (Windows, Linux), next to an option that turns the automatic checks off; ANNZARRO_DISABLE_UPDATES=1 turns updates off entirely. macOS, the Windows installer and the Linux AppImage update in place; a .deb and the Windows portable zip show a link to the new release. Checks contact only github.com. Updating starts with 0.5.3: users of 0.5.2 and older update by hand once (details).

  • Homebrew on macOS: brew install --cask settylab/annzarro/annzarro; the app then updates itself. The tap follows each release.

  • No title bar on macOS and Windows: the AnnZarro header is the title bar, at the same height as in the browser. On macOS the three window buttons sit at the top left and the AnnZarro title moves down below them. On Windows the minimise, maximise and close buttons, in the header’s colours in the light and the dark theme, sit at the top right, above the panel set buttons. Drag the empty part of the header to move the window and double-click it to zoom it. Linux keeps its normal window frame.

  • A File menu on every platform: Open Dataset… (Cmd/Ctrl+O, the dataset browser), Open File or Folder… (Cmd+Shift+O on macOS; Open .h5ad File… and Open .zarr Folder… on Windows and Linux), Open Recent and Refresh Dataset. The same buttons are in the dataset browser and on the welcome screen.

  • Drag a .zarr folder or an .h5ad file onto the window to open it. Several items open the first and say so; anything else is refused with a message.

  • .h5ad files open by double click and Open With, and .zarr folders through Open With or a drop on the Dock icon (on macOS they stay folders in Finder). A path on the command line opens in the running window. Opened datasets go to the Dock’s and jump list’s recent items, and the macOS window shows the file.

Admin: remote datasets

  • Whether the server may read s3://, gs://, gcs:// and http(s):// datasets is set in the admin interface: Data sources > Remote datasets (the section was called Data directories) offers Off, Only these locations (a validated list of URL prefixes) and Any location (with a warning that the server then fetches any URL, internal addresses included). It takes effect at once in every worker, is audited, and the dataset browser’s Remote place follows it. server.remote_stores, server.remote_allowlist and their environment variables stay the upper limit: an admin narrows them, never widens them, and the card shows what set the value.

  • The Settings switch “No external hosts” is renamed “Gene Set Analysis: no external services” and says it does not affect remote datasets. Its key and API name are unchanged.

  • The refusal “Remote datasets are disabled on this server” names Admin > Data sources > Remote datasets as well as server.remote_stores.

Changed

  • Plot Options opens with only the Export group unfolded; the groups you open or close are still remembered per panel.

  • The Save / Load / Close all / Share buttons in the header line up with the focus boxes instead of the middle of their labels.

  • Load Panel Set moved from Ctrl+O to Ctrl+Shift+L (Cmd+Shift+L); Ctrl+O opens the dataset browser. In the browser Ctrl+R reloads the page as the browser does and does not refresh the dataset (the Refresh dataset button does); the desktop app keeps Ctrl+R/Cmd+R for it.

  • The header wraps the panel set buttons below the focus boxes only when the focus boxes and the icon-only buttons do not fit one row; a dataset name never changes the header.

API

  • GET /api/v1/datasets/details?path=: a store’s counts, format, columns, obsm keys, layers and store-check findings, read from metadata only and stopped after server.dataset_details_check_budget_s (1.5 s by default).

  • sort= and order= on GET /api/v1/datasets/folder (name or modified, applied before paging), and folder= on GET /api/v1/datasets/search to search below one folder.

  • POST /api/v1/datasets/lookup answers which of a list of paths you may open, with the counts of each, and the same answer for a store that is not there and one you may not read. (HTTP API reference)

Docs

0.5.1 (2026-10-14)

A reworked Plot Options menu, with tick, margin, colour bar and scale bar controls, presets, and copying a plot’s look to other panels. The grey “Not in table” points follow their own count and have their own controls. Equal aspect plots keep their cells and zoom, and changing an axis on a large subset no longer freezes the page.

Plot Options

  • Plot Options is grouped into nine folding sections (Presets, Appearance, Axes and ticks, Points, Legend and colour bar, Text, Margins, Export, Apply to other panels). The menu opens again as you left it.

  • Publication preset: white background, no grid, five ticks per axis, larger text, a tighter margin. Default returns the look to a new plot’s; the font, the scale bar’s size and unit, the Points group and the export size stay.

  • Tick marks and tick labels switch separately, and Ticks per axis asks for about N ticks (2 to 30; empty is automatic).

  • Margins: one slider sets all four sides; Set each side separately shows a box for each.

  • Colour bar: its own position (right, left, top, bottom or with the legend), a length from 10 to 100 % and a tick count.

  • Scale bar (off by default) for embeddings and spatial coordinates: a bar of a round length in coordinate units, about a fifth of the width in view, redrawn as you zoom or resize, with a Size slider (length, line and label together), an optional unit text such as µm and a corner. It needs Equal aspect on a 2D plot: greyed out otherwise, and turning Equal aspect off switches it off. It is part of an export and not part of the Publication preset.

  • Apply to other panels copies a plot’s look (not what it shows) to every other cell plot or gene plot, after a confirm in the menu. All of it is kept in links and panel sets; links from before use the defaults.

Changed

  • Not in table: the grey points have no legend entry, like the highlighted cell. Plot Options has a “Not in table” group with their size (relative to the automatic size, 1x by default), opacity and colour; links, copies and saved panel sets keep them.

  • Without Equal aspect, switching the x field also resets a zoom on y, as reloading the page or opening a link already did.

  • Stable download links that always fetch the newest desktop app, for example https://github.com/settylab/annzarro/releases/latest/download/AnnZarro-macos-arm64.dmg. The README and the docs link their platform badges to them.

Fixes

  • With the grid off, tick labels no longer disappear on a plot’s first draw and reappear after the next edit: they follow Show Tick Labels alone.

  • Not in table: the grey points a table filter leaves in a plot are sized from their own count, the way the plot’s points are, instead of keeping the size they had before they were excluded. Setting the point size by hand after the table is active, switching the table on in the panel and opening a plot that names it now give the same grey points, and a different table or subset readjusts them.

  • Equal aspect: switching a 2D plot’s coordinates refits both axes together and keeps every cell in view. A zoom saved in a link from before v0.5.0 is kept only if it shows the data.

  • Equal aspect: a zoom is kept when the panel is resized, for example when a neighbouring panel is closed and the panel is split again.

  • Double-click (Reset axes) fits all the data again, also on a plot opened from a link with a saved zoom.

  • Changing a plot axis on a large subset coloured by a category no longer freezes the page. On a 100,000-cell subset of a 1.06 million cell dataset coloured by 60 categories, one axis change blocked the page for about 4 s and two in quick succession made the browser report “Page Unresponsive”; both now take 0.1 to 0.15 s.

  • Decoding a categorical column whose labels contain quotes or escaped characters takes 0.1 s instead of 3.8 s for 200,000 values with 20,000 such labels.

  • A slow reply to the page’s settings request no longer makes a large dataset download every cell name: the first dataset load waits for the settings (at most 10 s) before it decides whether names stay on the server.

Docs

0.5.0 (2026-10-13)

An admin web interface for shared servers: an administrator adds data directories, manages users, gives an external collaborator only some of the datasets, switches panels and external hosts off, and reads the server’s health and an audit log, all in the browser. Every request that changes something is protected against forged requests, the server resists requests that would hold a worker, and very large datasets are cheaper to open and read.

Admin interface

  • /admin (login on) has six sections: Users, Data directories, Server health, Panel sets, Settings and Audit log. A change needs the admin’s password entered within the last 30 minutes, and every change is written to the audit log.

  • Users: add and remove users, reset passwords (a generated password is shown once), grant or revoke admin, and clear a lockout. The last admin cannot be removed or demoted. Everyone can change their own password at Change password in the header.

  • Data directories: with server.admin_data_roots set, an admin adds directories in the browser, below the parents the operator allows. They appear in the dataset picker at once.

  • Dataset access for external collaborators: restrict a user to the folders, stores and remote prefixes you grant; everyone else keeps seeing everything. A restricted user gets the same “not found” for a dataset they may not read as for one that does not exist. Manage it in Users (Access) or with annzarro user access.

  • Settings: switch panel types off for everyone or for one user, and control what Gene Set Analysis may contact (all external hosts, individual services, the species lookup, the links). The browser refuses any host the settings do not allow.

  • Server health: each worker’s read budget in use, cache size, requests and refusals by reason, and who was seen recently. Panel sets: list by owner, reassign or delete.

  • Audit log: each record carries the hash of the one before it, so an edit in the middle is found (annzarro admin audit verify). The log also goes to the annzarro.audit logger.

  • A change made in the browser or with annzarro user ... reaches every gunicorn worker on its next request, without a restart.

Security

  • Every request that changes something needs the session’s token and a same-origin Origin. Sign-in has its own one-time token, and signing out is a POST.

  • Pages carry a Content-Security-Policy that lists the external hosts the user’s settings allow.

  • A request body over server.max_request_mb (64) is refused with 413.

Performance and limits

  • Cell and gene names are streamed. For a 95.6 million cell store the server holds 0.42 GiB and answers in 0.14 s, where it held 15.2 GiB and took 14.4 s.

  • Sharded stores: a 1,000-row subset of a 5M x 50 store in one 850 MB shard peaks at 47 MiB instead of 757 MiB.

  • Colouring by a gene on a subset of a CSR X reads each chunk once: 10,000 cells of a 1 million cell store take 0.4 s instead of 57 s, at about 200 MB more peak memory (#108).

  • Subsets share the cache and the compute fairly between users. A subset that cannot start within server.subset_wait_s (30 s) is answered 503 read_busy.

  • Numeric subset filters: six at once on a 95.6 million cell store take 1.8 s and 10.8 GB instead of 18.8 s and 30.5 GB.

  • Name search with a regular expression: a pattern that could hold a worker is refused, and every search stops after server.name_search_timeout_s (10 s). The picker says when a search ran out of time.

  • Remote stores: one request may spend at most server.remote_request_timeout_s on remote reads (120 s on a shared server, 600 s for one user), then fails with 504 remote_deadline.

  • Other limits: panel-set quotas per user and per server, a cap on the entries a directory listing scans, uns members sized before they are read, and gunicorn workers replaced after 20,000 requests. See Resource limits in the docs.

  • annzarro store check reports, from metadata only, what in a store’s layout makes AnnZarro slow, with the cost in clicks and a fix for each.

  • In large-plot mode, hovering finds the nearest point in 1 ms instead of 43 ms at 95.6 million points and skips categories hidden in the legend.

Fixes

  • Switching a plot axis between a text field and a number field draws and refits the axis for the new field. In large-plot mode the x and y pickers grey out categorical columns, with the reason, until a subset is made.

  • A table column named like orig.ident shows its values instead of N/A.

  • A 3D plot keeps its wheel zoom when the focused cell changes or the plot is redrawn.

  • The open dataset list follows the window height, with one scrollbar.

  • Center at 0 with Lock Range centres the colour scale on the locked range, and Log with Lock Range keeps the typed limits; turning either off restores them.

  • A panel opened from the welcome page before the first dataset loaded no longer keeps its “No dataset loaded” screen.

  • Splitting a panel, choosing a panel type and closing a section ease to the new sizes instead of jumping (not with the system’s reduced-motion setting).

  • A missing store is answered 404 not_found on every data route.

Desktop app

  • The macOS app is signed with a Developer ID and notarized by Apple. Open the .dmg and drag AnnZarro to Applications.

Docs

  • A Python API reference; pages on choosing the desktop app or a server, the admin interface, the Admin API, the policy registry and resource limits; CONTRIBUTING.md and a code of conduct.

  • The README and CITATION.cff cite the bioRxiv preprint.

Upgrade notes

  • Scripts that sign in with a cookie and send POST, PUT or DELETE must send the session’s token in X-AnnZarro-CSRF (the page has it in <meta name="annzarro-csrf">). curl and annzarro export against a server with login off are not affected.

  • Behind a proxy that terminates TLS or rewrites Host, set server.proxy_count, or every change is refused with 403 csrf_failed and the admin page refuses changes.

  • Sign-in and sign-out: a script that posts to /login must fetch the form first for its one-time login_csrf field. /logout is POST only.

  • Content-Security-Policy: a customised page or an embedding that contacts hosts the policy does not list is refused by the browser.

  • External hosts: Enrichr and Reactome are off by default on a shared server; switch them on in Settings or with integrations.gene_set.services. integrations.external_requests accepts none.

  • /status and /cache/info answer admins only on a shared server.

  • Remote deadline: keep the proxy’s proxy_read_timeout at or above server.remote_request_timeout_s (120 s); 0 turns the deadline off.

  • Subsets: raise server.subset_budget_mb and server.subset_wait_s if users see 503 read_busy.

  • Names streaming behind nginx: use proxy_http_version 1.1.

  • Admin state lives in annzarro_admin/ beside the users file; back it up with the users file. annzarro user commands must run as the account that owns that directory.

0.4.3 (2026-10-11)

Zarr format 3 stores, sharded ones included, are read correctly. A new dataset picker browses folders and searches, dataset names are never cut where you need them whole, large pairwise matrices are read safely, and stores on network filesystems open faster.

Zarr format 3

  • Sharded stores are read correctly, and the read limits size a sharded array by its inner chunk, not by the whole shard. With anndata 0.13’s default 1 GB shards, one row of a pairwise matrix used to count as 950 MB when the real read is about 10 MB.

  • A server running zarr 2 answers a format 3 store with a clear 400 unsupported_type instead of server errors or empty answers.

  • The remote chunk cache works for sharded stores. A remote server that ignores Range requests gets an error that names the cause.

  • zarr 3.0.0 to 3.0.8 are excluded: they read sharded arrays wrongly or lose members of consolidated format 2 stores. Use zarr 3.0.9 or newer, or zarr 2.

  • The docs say which server reads which format, list the measured costs, and give sharding guidance. Format 2 stays fully supported.

Opening datasets

  • A dataset picker opens from Browse… in the dataset list. Search by name and path (with a regex toggle), browse folders, see cells x genes, format and date for each store, find recently opened datasets at the top, and see why an unreadable store can’t be opened.

  • The dataset list shows full names; the dataset box cuts long names in the middle, keeping the end, and its tooltip holds the path.

  • The header stays on one row as long as it can: button labels collapse to icons first, then the focus boxes and the dataset box narrow.

  • The focused gene and cell boxes grow to show their whole value.

  • The typed-path field says what you can open there.

Large data

  • Pairwise matrices (obsp, varp) are read by rows, at most 1,000 per request, under a per-request memory limit (server.max_read_mb) and a server-wide read budget. A request that would decompress an n x n block is refused with a reason.

  • The first open on a network filesystem takes about 12 filesystem calls instead of about 15,000.

  • Stepping through the parts of a subset reloads only the panels that read cells.

  • On a plot over the cell limit, the notice says what is unavailable and how to get it back.

Plots

  • A colour column the dataset lacks draws grey with a “not in this dataset” notice, also on plots over the cell limit, which showed an error.

  • A gene plot with text coordinates (for example Ensembl IDs) counts every drawn gene as shown.

  • A default colour from one dataset is not carried over to a dataset that lacks it.

  • After a part step on a large plot, the strip no longer says “Plot went blank”.

Docs and data

  • One demo store for every doc, tutorial and panel set, bm_aging_annzarro.zarr, on Zenodo: 10.5281/zenodo.23277103.

Upgrade notes

  • zarr 3.0.0 to 3.0.8 are no longer accepted; upgrade to zarr 3.0.9 or newer.

  • obsp and varp reads are limited by server.max_read_mb; raise it if a request you need is refused.

0.4.2 (2026-10-08)

Very large datasets on shared servers: a billion-cell store steps through its parts and searches its cells without the server’s memory growing, plus fixes to clicks, point size, subsets and links.

Very large datasets

  • A subset with parts steps to the next part without closing the panel. A newer step replaces the one still loading, a “Loading part N…” pill shows progress, and the next part is read ahead. On a 1 billion cell store a step went from 58 to 72 s to about 23 s.

  • Above 5 million cells, cell names are loaded only when a view needs them (ui.defaults.names_on_demand_above).

  • Name search is bounded by memory (server.name_index_max_mb). Clicking the Focused Cell box on a 1 billion cell store grew the server by about 32 GB; it now stays near 1 GB.

  • The first plot of a 95.6 million cell store, and subsets of a 1 billion cell one, no longer wait for every name to be hashed.

  • A large plot that would not fit in the browser’s memory is declined with a suggested subset size instead of crashing the tab.

  • Large-plot mode has hover, click and the focused-cell highlight.

  • A Cancel button appears on a panel whose load takes longer than 1.5 s, and changing what a panel shows cancels its earlier load.

Plots

  • A cell is selected only by a short, still click, not at the start of a drag or a 3D rotation.

  • Automatic point size follows the points left after filters and subsets.

  • Boolean obs columns are coloured in the order False, True, as scanpy does, so stored uns/<col>_colors apply to the right values.

  • Applying a subset updates every affected panel.

Docs and data

  • The spatial demo ships a 5-step expression walk on the kNN graph (obsp/expression_walk_t5).

Upgrade notes

  • On Linux systems older than glibc 2.28 (for example Ubuntu 18.04), install with pip install "h5py<3.16" annzarro.