Skip to content
Sign inRequest access

Storage & media

Every file uMarker persists (screenshots and their thumbnails, widget logos, avatars, support-ticket images) is read and written only through a Laravel filesystem disk, never a raw filesystem path or a vendor SDK. That single seam is a founding principle: media is always tenant-scoped and access-mediated rather than reached by a raw path or a public URL.

Media is written under stable prefixes on the configured disk, each file named with a random UUID, so filenames are unguessable and never collide:

  • Directoryfeedback/
    • Directoryscreenshots/ the report screenshot, and its small WebP thumbnail beside it
      • …
    • Directoryattachments/ historical only: the widget no longer collects attachments
      • …
    • Directoryrecordings/ historical only: screen recording was retired
      • …

The database stores only the path to each file, not the bytes. Structured capture data (console logs, metadata, breadcrumbs, failed requests, pin geometry) lives as JSON columns in MySQL, not on the disk.

Before the upload: sized and encoded on the device

Section titled “Before the upload: sized and encoded on the device”

The screenshot is prepared in the reporter’s browser, by one shared encoder used by the snippet, the annotator and the extension, because two copies of an encoder is how one surface quietly stores twice what the other does.

  • Pixel budget. A render larger than 8 megapixels is scaled down to fit, floored at half scale, because an unreadable screenshot is not a cheaper screenshot.
  • WebP, checked rather than assumed. The canvas is encoded as WebP at quality 0.85 and the result is verified to actually be WebP, because a browser handed a type it doesn’t support silently returns a PNG, which would be several times larger than the JPEG it was trying to beat. Browsers without WebP encoding fall back to JPEG.

Storage is a trust boundary. When the ingest endpoint decodes an uploaded data URL, it does not trust what the client says the file is:

  1. Size is checked before decoding. The base64 length is bounded first (a denial-of-service guard), then the decoded byte length is checked against the cap.

  2. Images must really be images. A screenshot is decoded and inspected: it must be a genuine raster image, and the stored MIME type is the detected type, not the client-declared one. This rejects HTML- or SVG-as-PNG spoofing that could otherwise become stored XSS when the image is served inline.

  3. Only known types are stored. The extension on disk is chosen from one allowlist keyed on the detected MIME, and that same table is read back to tell the gallery what a stored file is, so the two directions cannot drift.

  4. A refused write fails loudly. The disk layer raises rather than returning false, because the silent version stored a path to an object that was never written: during a storage incident every report would be accepted with a 200 and show a broken image for ever, un-recapturable because the visitor is long gone.

Media Cap
Report screenshot 12 MB decoded
Support-ticket image 5 MB decoded, images only, at most 3 per message

Every screenshot’s byte size is recorded on the report, and each account carries a running storage_bytes counter. The counter is treated as a cache, not as truth: bulk deletes bypass model events and jobs fail, so a counter that can only ever be adjusted is eventually wrong. A nightly storage:reconcile sums the rows and corrects it.

What the meter covers is the media that scales with usage: report screenshots, the attachments and recordings older reports still carry, and the images you attach to a support ticket. Images support sends you are our cost, not yours, and are never counted. Widget logos and announcement images are one small file per project or per post; counting them would make the number harder to explain without making it more accurate.

Two scheduled jobs delete media, and the distinction between them is deliberate:

Your retention policy

feedback:purge-expired runs nightly and hard-deletes whole reports older than the window your account set, media included. This is a privacy promise you make, so it only runs for accounts that set a window and whose plan includes the custom-retention capability.

The plan's media window

media:purge-expired clears only the bytes past your plan’s media window and keeps the report: its title, description, pins, comments, labels, status and technical context all survive. Free 30 days, Starter 180, Professional 365, Agency 730.

A weekly sweep also removes files that no row claims any more, because a failed deletion job leaves bytes nobody can find from a row and nobody stops paying for.

Reading a file back is authorized, tenant-scoped, and hardened against content-type tricks:

  • Tenant-scoped binding. A screenshot resolves through an account-scoped model, so a request for another tenant’s file returns 404 rather than the bytes, and the reader must also be able to reach the project the report belongs to.
  • Downloading is its own right. On the endpoints that authenticate with your session — the full screenshot on the report page, a recording, an attachment — serving the file requires the separate feedback.download permission beyond the right to read the report. A screenshot is a photograph of someone’s screen mid-session, the one artefact in this product most likely to hold something nobody meant to send. That is why reading a report and taking its files away were split apart. The signed links below are a second door with its own rule: there the signature is the whole credential, and that is how the inbox board’s thumbnails and the gallery are served.
  • The plan’s history window applies. A report outside your plan’s visible history does not serve its media through the dashboard either. A signed link already issued runs on the media clock instead, and the two can differ sharply: on Free, history is 7 days and the media window 30.
  • No MIME sniffing. Every media response sets X-Content-Type-Options: nosniff, so a browser won’t reinterpret a file as a different, executable type.
  • Signed links for outside readers. A tracker ticket — GitHub, Linear or Jira — carries the screenshot as a signed URL, because the person opening it there has no session with us. Its expiry is computed from the report’s own media-retention clock, so a link never claims more life than the file has. A report already past its window produces no link at all rather than one that 404s. Slack, Teams, Telegram and webhook messages carry no image at all: only a link back to the report in uMarker, which needs a session.
  • The hosting region is not announced. It is a deployment setting and has not been decided and published. Nothing on this page should be read as a residency commitment.
  • Backups hold deleted bytes longer than you do. Database backups are kept for 30 days by default, so a purged report survives inside them until that window passes. A retention promise that ignores backups is false.
  • If the media bucket has object versioning switched on, a deleted screenshot remains as a non-current version until a lifecycle rule expires it. That rule is a bucket setting rather than application code, and it must be in place for the retention promise to hold end to end.