A digital file outlives its creator’s software more often than people expect. A spreadsheet from 1992 or a word processor file from 1987 can be perfectly intact, every byte present, and still be hard to open, because the program that understood it is gone and its format was never written down. Choosing preservation file formats is a way of making that problem smaller before it starts.
The choice is never between a perfect format and a bad one. It is a set of trade-offs between file size, fidelity, convenience and risk, made differently for a photograph, a thesis and a database. This page walks through the criteria, then through the main content types, with the compromises named.
The sustainability factors
The most widely cited framework comes from the Library of Congress, whose format descriptions judge every format against seven sustainability factors. They are worth knowing by name, because most format policies borrow from them.
Disclosure. Is there a complete, public specification, and are tools available to check files against it? A format whose documentation is incomplete, or available only under a non-disclosure agreement, depends on its owner staying interested. Formats with only part of the specification published sit in an awkward middle ground called partial disclosure.
Adoption. How many people and organizations use the format, and how many independent programs read and write it? Wide adoption creates commercial and community pressure to keep a format readable, and it multiplies the number of working implementations someone can study later.
Transparency. How easily can the file be analyzed with basic tools, including human reading? Plain text and uncompressed images score high. Heavily compressed or encrypted formats score low, because one misunderstood detail can make the whole file unreadable.
Self-documentation. Can the file carry its own descriptive and technical metadata? A TIFF with embedded capture information or a Broadcast WAV with its origination data is easier to understand decades later than a bare stream of samples.
External dependencies. Does rendering the file need particular hardware, software, fonts, plug-ins or network resources? Every dependency is something else that has to survive.
Impact of patents. Patents can restrict who may build tools for a format, which slows open-source support and can make future implementations expensive. Most patents expire after twenty years, so this factor changes over time.
Technical protection mechanisms. Encryption, digital rights management and password protection block the access that preservation depends on. Archives generally refuse protected files or remove the protection on ingest where they have the right to.
The Library of Congress also publishes a Recommended Formats Statement, reviewed every year, which turns these factors into preferred and acceptable lists for each kind of content. It is the most practical single reference for anyone writing a policy.
Still images
For digitized photographs, maps and documents, uncompressed TIFF remains the default master format in most cultural heritage institutions. The baseline TIFF 6.0 specification has been stable since 1992, almost every imaging program supports it, and an uncompressed file is easy to inspect and repair. The cost is size: a large-format scan can run to hundreds of megabytes. LZW-compressed TIFF cuts that losslessly, and the LZW patents that once made it controversial have long expired.
PNG (ISO/IEC 15948) is lossless, open and universally supported by browsers, and suits born-digital graphics, screenshots and diagrams. It has no CMYK mode and limited support for the rich technical metadata that photographic workflows rely on, which is why it is seen less often as a digitization master.
JPEG 2000 (ISO/IEC 15444) has a reversible, lossless mode and much better compression than TIFF, and several large libraries use it for mass digitization. The trade-off is software support: fewer programs read it well, and the specification is complex, which lowers transparency.
Born-digital JPEG photographs are a separate case. Converting a camera JPEG to TIFF makes the file larger without restoring anything lost in compression, so the usual advice is to keep the original. Camera raw files are kept where they exist; DNG is a documented alternative to proprietary raw formats.
Documents
PDF/A (ISO 19005) is the archival profile of PDF. It forbids or restricts features that threaten long-term rendering: all fonts must be embedded, encryption is not allowed, and external content and scripts are excluded. It comes in four parts:
- PDF/A-1 (2005), based on PDF 1.4, with conformance levels A (accessible, tagged) and B (basic visual appearance).
- PDF/A-2 (2011), based on PDF 1.7, adding JPEG 2000 compression, transparency, layers and embedded PDF/A files.
- PDF/A-3 (2012), identical to part 2 except that any file type may be embedded, for example the spreadsheet a report was built from.
- PDF/A-4 (2020), based on PDF 2.0, with simplified conformance levels.
A file that calls itself PDF/A in its metadata is not necessarily compliant. Validation with a tool such as veraPDF is part of the job.
PDF/A preserves appearance well and editability poorly. For documents that may need to be edited, reused or mined for text, OpenDocument (ISO/IEC 26300) is the open, XML-based alternative. Office Open XML (ISO/IEC 29500) is more widely used in practice and is generally accepted, though the specification is very large.
For text where layout does not matter, plain text in UTF-8 and XML with a published schema remain the most transparent options there are. A TEI-encoded edition of a manuscript, for example, is readable with nothing more than a text editor.
Audio
WAV with linear PCM audio is the standard master format for digitized sound. Broadcast WAV (EBU Tech 3285) adds a bext chunk for origination metadata such as the source, the date and the coding history, which is why sound archives favor it. Standard WAV files are limited to 4 GB because of a 32-bit size field; RF64 and BW64 extend that for long multi-channel recordings.
FLAC is lossless, open and now documented as RFC 9639. It roughly halves file size compared with WAV, and many archives accept it as a preservation format. A common compromise keeps WAV as the master and FLAC for distribution, or FLAC as the master where storage is the main constraint.
MP3 and AAC are lossy. They are fine as access copies and should be kept as received when they are the original, but no one should create a new master in them.
Moving image
Video is where the trade-offs are hardest, because uncompressed files are enormous: an hour of uncompressed standard-definition video occupies on the order of 100 GB, high definition several times that.
FFV1 in a Matroska container has become a widely used answer for digitized analog video. FFV1 is a lossless codec, specified in RFC 9043; Matroska is an open container, specified in RFC 9559. Both were standardized through the IETF CELLAR working group, are free of patent licensing, and are supported by open-source tools such as FFmpeg. FFV1 also stores a checksum for each slice of a frame, so damage can be located precisely. Archivematica, an open-source preservation system, uses FFV1 in Matroska as its default video normalization target.
Delivery codecs are a different matter. H.264 in MP4 is the most common video format in circulation, readable on almost every device, and that adoption counts for a lot. But it is lossy, it relies on patent licensing, and it is designed for viewing, not for faithful reproduction of the source. An archive that receives born-digital H.264 files usually keeps them as they are. An archive digitizing tape should not choose H.264 as the master.
Data, web and email
Tabular data is often best kept as CSV, with its limitations handled in documentation. RFC 4180 describes a common form of CSV, but real files vary in delimiter, quoting, line endings and encoding. A preservation-ready CSV states its encoding (UTF-8 by preference) and comes with a data dictionary explaining each column and its units. Spreadsheets kept in XLSX or ODS also carry formulas and formatting, which CSV loses.
Databases are harder. SQLite is one of the few database formats with a stable, fully documented file structure, and the Library of Congress lists it as a recommended storage format for datasets. Larger relational databases are usually exported to SQL dumps or to an archival package such as SIARD.
XML and JSON work well for structured data, provided the schema (XML Schema, JSON Schema or a written specification) is kept alongside them.
Web archives use WARC (ISO 28500), which stores the HTTP requests and responses captured by a crawler along with metadata about the crawl. It is the format used by most web archiving programs and by tools such as Heritrix and Webrecorder.
Email is commonly kept as EML (one message per file, following RFC 5322) or MBOX (many messages concatenated in one file). MBOX has several incompatible variants, so the variant should be recorded. Attachments travel inside the messages in their own formats and need identification of their own.
Summary table
| Content type | Preferred | Acceptable |
|---|---|---|
| Still images | Uncompressed or LZW TIFF; PNG for born-digital graphics | JPEG 2000 lossless; original JPEG or raw as received |
| Text documents | PDF/A; ODF; plain text or XML with schema | Office Open XML; ordinary PDF as received |
| Audio | WAV or Broadcast WAV (linear PCM) | FLAC; original MP3 or AAC as received |
| Video | FFV1 in Matroska; uncompressed in MOV or MXF | Original H.264 or other delivery formats as received |
| Data | CSV in UTF-8 with documentation; SQLite; XML or JSON with schema | XLSX, ODS; SQL dumps |
| Web and email | WARC; EML or MBOX | Other export formats with documentation |
Normalization and keeping the original
Normalization means converting incoming files into a small set of preferred formats, usually at ingest. It reduces the number of formats an archive has to monitor and support, which is a real gain for a small team.
It also introduces risk. A conversion can drop embedded metadata, change color profiles, flatten layers, lose formulas or alter layout, and the result may look correct while differing in ways no one checks. For that reason the dominant practice is to keep the original bitstream alongside the normalized copy, and to record the conversion tool, its version, its settings and the date. When a better tool appears, the original can be converted again.
Whether a conversion is lossless matters here too. Format registries describe compression as lossless or lossy (compression lossiness); a lossless-to-lossy conversion should never be the only copy kept.
Checksums and fixity
A format choice protects against obsolescence. It does nothing against bit rot, failed transfers or accidental edits. That is the job of fixity checking: computing a checksum for each file when it arrives, storing it, and recomputing it at intervals to confirm the file has not changed.
MD5 and SHA-1 remain common for fixity because accidental corruption is what they are guarding against, and for that they still work. SHA-256 and other members of the SHA-2 family, such as SHA-224, are preferred where deliberate tampering is a concern. Whatever the algorithm, it must be recorded with the value, and checks are only useful if there is a second copy to restore from when one fails.
Writing a file format policy
A file format policy turns all of the above into rules. A typical policy names, for each content type, the formats the archive prefers, the formats it accepts, and what happens to everything else: accepted with bit-level preservation only, converted on ingest, or refused. Many policies attach a level of commitment to each tier, so that a depositor knows that a TIFF will be actively migrated if needed while an obscure CAD file will be kept as received and nothing more.
The format itself is recorded in the descriptive metadata, typically as a media type in dc:format, with the format modeled as a dcterms:FileFormat or a dcterms:MediaType. Identification has to come first, though: a policy that prefers PDF/A is only enforceable if the archive can tell a PDF/A file from an ordinary PDF, which is a task for the tools covered in the file format identification guide. The hard cases are seldom the formats on the preferred list. They are the files nobody has identified yet, sitting in a deposit folder named misc.
Questions and answers
What makes a file format suitable for long-term preservation?
An open, complete specification; wide adoption with several independent implementations; a structure simple enough to analyze; the ability to carry its own metadata; few external dependencies; no blocking patents; and no encryption or rights management. The Library of Congress groups these as its sustainability factors.
Is PDF a preservation format?
Ordinary PDF can contain features that undermine long-term rendering, such as unembedded fonts, external links, scripts or encryption. PDF/A (ISO 19005) is the archival subset that forbids or restricts those features. A PDF/A file still needs to be validated, for example with veraPDF, before the claim is trusted.
Which PDF/A part should be used?
PDF/A-2 is a common default for new documents because it allows JPEG 2000 compression, transparency and embedded PDF/A files. PDF/A-3 allows any embedded file, useful for attaching source data. PDF/A-4, based on PDF 2.0, is the current part. PDF/A-1 suits systems that only accept the oldest profile.
Is JPEG acceptable for preservation?
For photographs that were born as JPEG, yes: keeping the original camera file is better than converting it, which cannot recover lost detail. For new digitization masters, JPEG is usually only acceptable as an access copy, because its lossy compression discards information every time the image is saved.
Why is uncompressed TIFF still recommended?
Baseline TIFF is extremely well documented, supported by almost every imaging program and easy to repair at the byte level. Uncompressed files are large, but a single damaged byte affects only a few pixels. LZW compression saves space losslessly and is widely accepted; its patents expired years ago.
What is the best preservation format for video?
For digitized analog video, FFV1 in a Matroska container is widely used: it is lossless, openly specified (RFC 9043 and RFC 9559) and free of patent licensing. Uncompressed video in MOV or MXF is an alternative with much larger files. Born-digital H.264 files are usually kept as received.
Is CSV a good preservation format for data?
CSV is simple and readable by almost any software, which is its strength. Its weakness is that it records no encoding, delimiter, data types or column meanings. It works well for preservation when paired with documentation: the character encoding (ideally UTF-8), a data dictionary and a note on the dialect used.
Should an archive normalize all files to preservation formats?
Normalization makes a collection easier to manage but every conversion risks losing something. The common approach keeps the original bitstream and adds a normalized copy, with a record of the tool, settings and date. Some archives only normalize when a format becomes hard to render.
How often should fixity be checked?
There is no single correct interval. Many repositories check on ingest, after every transfer or migration, and on a regular schedule such as yearly or quarterly for stored copies. What matters is that the checksum algorithm and the results are recorded so that a failure can be traced and repaired from another copy.
What is a file format policy?
A file format policy is a written statement of which formats an archive accepts, which it prefers, what it does with the rest, and what level of preservation it promises for each. It makes acquisition decisions consistent and tells depositors what to send.