Cool LUIs Dont Change

After my last note on identifiers, Leo Talirz pointed me to a great riff1 on Tim Berners-Lee’s classic note2 on “cool URIs”.

In the “Cool DOIs” article, Fenner breaks down a DOI into three parts: proxy, prefix, and suffix. A proxy is a server that maintains a map from prefixes to registrants. Example proxies are and An example prefix is 10.5281. and thus should return the same information: that given a DOI with prefix 10.5281, e.g. 10.5281/ZENODO.31780, is the registrant from which you can resolve the full DOI. Thus, when you ask for, the proxy looks up 10.5281 and tells your web client to ask for a URL corresponding to 10.5281/ZENODO.31780. The URL is at, meaning folks at registered the DOI with

McMurry et al.3 characterize a URI-as-identifier in a similar manner: as resolver, prefix, and local ID. A resolver can be a “primary” resolver, e.g., but it can also be a so-called “meta-resolver”4, e.g. or You register a prefix with a meta-resolver, and you also register resolution providers for your prefix. For example, someone registered the doi prefix with, along with a resolution provider with URL pattern{$id}. Because and share registrations, you can ask for, which meta-resolves to via URL pattern filling, and takes it from there. Another registered prefix is uniprot, with provider patterns{$id} and{$id} (so a meta-resolver can try an alternative if the primary provider is down). or yield the same result. uniprot:A0A022YWF9 is an example of a so-called compact URI (CURIE).

With meta-resolvers, you have semantic flexibility in your choice of prefix. Fenner emphasizes that DOI prefixes should be random and opaque because registrant/organization names can change. With meta-resolution, if UniProt changes their name, they can register a new prefix and encourage its use while still supporting the uniprot prefix. However, local IDs should be unique. My title for this note reflects this revision: cool local unique identifiers (“LUIs”4) don’t change.

Berners-Lee’s note gives sage advice related to rolling your own resolver for LUIs via a server hosted at a domain name you control. For semantically flexible prefixing, qualify everything with creation date. The precision of this date can reflect an acceptable cadence for updates: for research data projects, I think month-level precision, e.g. /YYYY/MM/, is acceptable. Thus, https://<domain>/2021/06/<org>/<LUI> reads as “ask <domain> for the record for <LUI> under the namepace /2021/06/<org>, i.e. from the data repository that, in June 2021, <domain> knew by the name <org>.

You can register these https://<domain>/YYYY/MM/<org>/ namespaces as prefixes with meta-resolvers and/or within your data products (e.g. as prefixes in RDF serializations). For a given /YYYY/MM/ qualification, your organization <org> can reflect the semantic partitioning strategy du jour (or rather, du mois), e.g. {/YYYY/MM/mp/calculations/, /YYYY/MM/mp/materials/, /YYYY/MM/mp/structures/,…}. The important thing here is that each such prefix is merely an alias for a permanent and semantically opaque repository ID within <domain>, akin to the 10.5281 example for DOIs.

This leaves us, at last, to LUIs. For these, I agree with Fenner’s note and with Geewax5: use random integers encoded with Crockford’s Base32 and a checksum. For example, using the base32-lib Python packaged by Invenio (spun out of CERN) folks:

import base32_lib as base32

id_encoded = base32.generate(
print(id_encoded)  # tw0t-ywdj-94
id_decoded = base32.decode(
print(id_decoded)  # 923446243762
id_encoded2 = base32.encode(
print(id_encoded2)  # tw0t-ywdj-94 

In the example, I specify the total length of encoded strings to be 10 characters, including 2 characters for the (ISO 7064, MOD 97-10) checksum. Thus, strings decode to 40-bit integers – 8 characters * (log2(32) = 5) bits/character. There are 2**40 ~ 1 trillion possible LUIs.

What’s nice about base32 encoding is that you can insert optional dashes anywhere for readability. Here I insert one every 4 characters, yielding LUIs like tw0t-ywdj-94. The encoding is also case-insensitive, and excludes the letters I, L, and O, because they can be visibly confused with the numbers 1 and 0 – when decoding, a supplied letter O will be replaced with the number 0. The letter U is also excluded to avoid accidental obscenity.

You can use up to 12 characters (14 with checksum) for encoding up to ~ 1.1 quintillion (10**18) LUIs and still store them internally as 64-bit integers. I think that my 8-character (12 with checksum and dashes) example above is a good tradeoff for compactness and cardinality – you can always create another repository under your domain, e.g. for transient/internal LUIs.

Finally, I do recommend recording your LUIs in an index so that collisions during generation, while unlikely, are thwarted. Being able to encode LUIs as integers can help ensure fast index lookups.

Subscribe to get short notes like this on Machine-Centric Science delivered to your email.

  1. M. Fenner, “Cool DOI’s,” DataCite Blog, 2016. (accessed Jun. 15, 2021). ↩︎

  2. T. Berners-Lee, “Cool URIs don’t change.,” 1998. (accessed Jun. 15, 2021). ↩︎

  3. J. A. McMurry et al., “Identifiers for the 21st century: How to design, provision, and reuse persistent identifiers to maximize utility and impact of life science data,” PLoS Biol, vol. 15, no. 6, p. e2001414, Jun. 2017, doi: 10/b88j↩︎

  4. S. M. Wimalaratne et al., “Uniform resolution of compact identifiers for biomedical data,” Sci Data, vol. 5, no. 1, Art. no. 1, May 2018, doi: 10/gdh496↩︎ ↩︎

  5. J. J. Geewax, API Design Patterns. O’Reilly Media, 2021. ↩︎