Skip to content

Survivorship rules that decide which value wins on a merged attendee record

Unified dataUpdated 2026-08-238 min read

In short

Survivorship rules decide which value wins for each attribute of a merged record. Running them one attribute at a time lets the name come from one source row and the phone from another. Three rules cover most registration fields: prefer the most recent, prefer a corporate email, and prefer the longest name that is not a nickname.

The deduplication run finished, three registration rows were correctly judged to be one person, and the record that came out of it names her K Odell, has no phone number and no job title, and lists a Gmail address.

Every input was fine. The matcher was right that the three rows are one buyer. What went wrong is downstream of the match: survivorship ran at record level, the newest row won, and the newest row was a two field re-registration typed on a phone.

Survivorship rules are the part of an identity pipeline that decides which value wins, and they are usually written last, in a hurry, by whoever is closest to the merge code.

The merge decision does not choose a value

It is worth separating the two questions, because teams routinely answer the second by accident while answering the first.

The matcher answers whether two records describe one person. That question has an answer in the world, and it is either right or wrong. Which of two conflicting job titles belongs on the resulting record is a different kind of question, with no fact of the matter to appeal to when both were true at different times.

Bleiholder and Naumann's 2008 survey in ACM Computing Surveys sets out the options for handling that second question. Dong and Naumann's 2009 tutorial at the VLDB conference restates the same classification in three lines. Strategies that ignore conflicts "are not aware of conflicts, perform no resolution, and thus may produce inconsistent results". Strategies that avoid them are aware but make one blanket decision for every case. Strategies that resolve them "provide the means for individual fusion decisions for each conflict".

They add a second axis that matters more for registration data than it first appears. "Deciding strategies choose a preferred value among the existing values, while mediating strategies can produce an entirely new value, such as the average of a set of conflicting numbers."

Almost every survivorship rule you want on an attendee record is a deciding strategy. Mediating between two job titles produces something nobody holds, and averaging two phone numbers produces a number that rings nowhere. Keep mediation for genuinely numeric aggregates and pick an existing value everywhere else.

Run the rules once per attribute

The single decision that separates a usable merged record from the one at the top of this post is the level the rules run at.

Record level survivorship picks a winning row and copies it. It is one line of code, it is easy to explain, and it throws away every good value in the losing rows. Attribute level survivorship asks the question once per field, so the same merge can take the name from a 2023 row, the phone from a 2025 row and the email from two rows that agree.

The objection to attribute level work is that it can produce a combination that never existed in any source, and the objection is correct. A record showing a 2023 name with a 2026 job title describes a person who was never described that way by any single system. That is the intended behaviour, since the merged record is a current best view assembled from evidence, and the way to stay honest about it is to keep the source and timestamp for every surviving value, which is golden record construction in J27.

Three rules on three rows

Take the three source rows behind the broken record above.

The first is a web registration from March 2023. Name Kathryn Odell, email [email protected], no phone, job title Marketing Manager, company Verdant Films Ltd.

The second is a lead capture scan from September 2025. Name Kate Odell, email [email protected], phone +447700900412, job title Head of Marketing, company Verdant Films.

The third is a web registration from February 2026. Name K Odell, email [email protected], no phone, no job title, company Verdant Films Ltd.

Now apply the rules field by field.

Name: longest form that is not a nickname. Kathryn Odell wins over Kate Odell and over the initial. The rule needs a nickname table, since Kate is a shorter form of Kathryn rather than a typing error, and without the table a plain longest string rule would still work here and would fail on a hyphenated surname typed in full on the sparser row.

Email: prefer corporate over consumer. The two corporate values agree, so the choice is easy, and the consumer address loses even though it appears on the row with the most complete data. Keep it as a secondary address rather than discarding it.

Phone: only one candidate. The 2025 row is the only source with a number, so it survives by default. This is the case record level survivorship gets wrong most often and most invisibly, because a null in the winning row silently becomes a null on the golden record.

Job title: most recent non empty value. The 2026 row is empty, so the rule falls through to the 2025 value, Head of Marketing. That fall through matters, and the version of the rule that takes the newest row's value including its blanks is the version that produces empty golden records.

Company: most recent, after normalisation. Verdant Films Ltd and Verdant Films are the same firm, so the choice is cosmetic once the names are normalised, and the legal suffix question belongs with legal entity suffixes in J18.

Five attributes, three sources, and the winning values come from all three rows. No single source row could have produced this record, which is the point of running the rules per attribute.

Which rule belongs on which field?

The mapping is worth writing down once and reviewing annually, because the defaults are wrong for several fields that matter.

Fields that describe a moment, such as job title, seniority, employer and phone, take the most recent non empty value. Fields that describe a person, such as first name, surname and date of birth where you hold it, take the most complete or most frequently observed value, since they do not change and a difference is usually an error. Fields that are graded, such as email, take a preference rule over a recency rule.

Two categories should never take a survivorship rule at all.

Consent and preference flags are the first. A marketing permission granted on one registration form is scoped to that form, that brand and that moment, and letting a 2026 opt in survive onto a record that also carries a 2024 opt out inverts the meaning of both. The correct behaviour is to hold every consent event separately with its source and timestamp, and to compute permission at send time. Anything else builds a compliance problem into the merge logic, and the legal framing of that sits with the privacy cluster rather than here.

Identifiers issued by other systems are the second. A CRM contact id, a badge number and a registration reference all belong to their source, and collapsing them onto one field loses the mapping you need to write results back.

What happens when two rules disagree?

Sometimes two rules apply to one field and point in different directions. The 2025 row has a corporate email that is older than a consumer one from 2026, so recency says one thing and preference says another.

Order the rules explicitly per field, rather than letting the order fall out of the code. Write it as a short ordered list in configuration, with preference ahead of recency for email and recency ahead of completeness for job title, and let a reviewer read the configuration without reading the implementation.

Where two candidate values remain after the ordered rules, the honest options are to keep both with a flag, or to send the field to review. Picking arbitrarily is defensible only if the choice is recorded as arbitrary, and ranking the sources themselves is a separate exercise, which is conflicting values across sources in J29.

Ties on timestamp are more common than they look, because a batch load stamps every row with the load time and two registrations taken on the same day arrive with the same date. A recency rule then has nothing to sort on and falls back to whatever order the database handed the rows over in, which changes between runs. The same three source rows can produce two different golden records on two consecutive loads, and the difference shows up as a phantom change in whatever report reads the field.

The fix is a deterministic final tie break, applied after every other rule: sort by source rank, then by source row identifier, and take the first. It resolves nothing intellectually and it makes the pipeline reproducible, which is what lets you compare this week's output against last week's and conclude that a difference means something. Store the timestamp at event granularity where the source gives you one, since a date alone throws away the ordering you actually had.

Where this stops

Survivorship rules are guesses with a policy attached, and two situations defeat them.

The first is the person who genuinely holds two current values. A consultant with two employers, a delegate with a work and a personal phone both in use, a director who is Head of Marketing at one subsidiary and a board member at another. The rules will pick one and present it as the value, and the loss is invisible on the golden record. Where the field matters commercially, model it as a set rather than a scalar and accept the extra cost in every query that reads it.

The second is a bad merge upstream. Survivorship applied to two records that are actually two people produces a coherent, well formed record describing nobody, and it does so with more confidence than the raw rows had. The rules amplify a matching error and tidy away the traces of it, which is an argument for keeping the source rows intact and the merge reversible, and for the recency rule in particular deserving the closer look it gets in the most recent wins rule in J28.

Write your current survivorship behaviour down this week, field by field, from the code rather than from memory, and mark every field where the rule is really pick whichever row the query returned first. On a typical pipeline that list is longer than anyone expects, and it is the shortest route into a unified data conversation with something concrete in it.

Questions people ask about survivorship rules

What are survivorship rules in data matching?
They are the rules that pick a surviving value once two or more records have been judged to describe the same person. Each attribute gets its own rule, so a merged record is assembled field by field from the source rows, and no single row is copied wholesale onto the result.
Should survivorship run per record or per attribute?
Per attribute. A single source row is rarely best at everything. The newest registration often has the freshest job title and the sparsest everything else, while an older row may hold the only phone number. Choosing one winning row discards good values that a per attribute rule would have kept.
Which survivorship rules cover most registration fields?
Three do most of the work on attendee data. Prefer the most recent non empty value for fields that change, such as job title. Prefer a corporate email address over a consumer one. Prefer the longest version of a name that is not a nickname, so Kathryn beats Kate and Kate beats an initial.

Related reading

All identity resolution articles