The most recent wins rule and the two fields it gets wrong
The most recent wins rule takes the value from the newest source row. It suits fields that genuinely change, such as job title and phone number. It fails on name and email, where the newest row is often a hurried on-site registration, so a completeness check and a source trust rank belong ahead of recency.
A delegate emails to ask why her badge says M Chenn. The record behind it holds a surname with a doubled letter, no phone number and no job title, and until March it held all three correctly.
What happened in March is that she came to the desk, the queue was long, a colleague typed her details into the kiosk, and the resulting row was newer than everything else in the file. The most recent wins rule did exactly what it says.
The rule is not wrong. It is the right rule for a specific class of field and the wrong rule applied everywhere else, and the cost of the difference shows up on badges.
When does the rule earn its place?
Some attributes change because the world changed. A delegate who was a Quality Systems Manager in 2021 and an Operations Director in 2026 has two true values, and the newer one is the one you want on a record used for targeting, segmentation and sales conversations.
Job title is the clearest case. Employer is the next. Mobile numbers, postal addresses and seniority all belong in the same group, since they describe a current state and a stale value is straightforwardly wrong rather than merely old.
Dong and Naumann make a related point in their 2009 VLDB tutorial when they observe that a value can be in a third case beyond true and false, which is out of date. That framing is useful in a registration archive, where a large share of the disagreements between two rows are neither errors nor contradictions, only the same fact observed at two different times.
Two rows, five years apart
Take the record behind the badge complaint.
The first row is a web registration from June 2021. First name Meixiang, surname Chen, email [email protected], phone +441615960117, job title Quality Systems Manager, company Northgate Instruments, country GB. Seven of eight fields populated.
The second is an on-site kiosk registration from March 2026. First name M, surname Chenn, the same email address, and nothing else. Three of eight fields populated.
Pure recency gives the 2026 row every field it has a value for, and gives nothing where it is empty, so the assembled record shows M Chenn with no phone and no job title. Completeness tells a different story: the older row is 88 per cent populated against 38 per cent for the newer one.
The right answer here is not obvious from either number on its own. The 2026 row is genuinely more recent and its email agrees with the older row, which is real evidence. Its surname is one keystroke away from the older one, which in a kiosk queue is much more likely to be a slip than a legal name change.
So measure the sources. Take 300 kiosk rows and count how many had a surname corrected afterwards, and do the same for 300 web self registrations. If 21 of the kiosk rows were corrected against 4 of the web rows, that is 7 per cent against 1.3 per cent, and the kiosk source is roughly five times as likely to introduce a name error. That ratio is the fact your rule should be built on, and it takes an afternoon to produce from data you already hold.
Recency is one axis of three
The research on this problem stopped treating freshness as the whole question some time ago. Dong, Berti-Equille and Srivastava, writing in the Proceedings of the VLDB Endowment in 2009, model the quality of sources over time using three measures together: coverage, exactness and freshness. Their point is that low quality data has several causes that deserve different handling, since some sources make errors, some provide correct data and fail to update it, and some update slowly.
Their study is worth knowing for its scale as well as its method. They collected data on Manhattan restaurants from 12 web sources weekly over three months, found 5,269 restaurants mentioned by at least two sources, and identified 280 that had recently gone out of business. Sources disagreed constantly, and the disagreements split into out of date values and outright errors, which a pure recency rule cannot distinguish.
Map those three measures onto a registration pipeline and you get a usable design. Coverage is how many fields a source populates, which you can compute per source per field. Exactness is its correction rate, which is the 7 per cent against 1.3 per cent above. Freshness is the timestamp. A survivorship rule that reads only the third of these is using one third of the information sitting in your own warehouse.
The two fields it gets wrong
Two attributes on a registration record should almost never be decided by recency alone.
The first is the person's name. Names change, through marriage and other routes, and those changes are rare enough that a difference between two rows is far more likely to be a shortened form, an initial or a typing error. A rule that takes the newest spelling will steadily degrade your name field towards whatever the fastest data entry route produces, which is initials and doubled letters. The alternative treatment sits with survivorship rules in J26, which sets out the full rule set per field.
The second is the email address. A delegate who registers on their phone with a personal address after four years of using a work one has given you a newer address and a weaker one, since the corporate address identifies their employer, is likely to be monitored during working hours, and is the one their colleagues would recognise. Recency here trades an identifying value for a convenient one.
Both failures share a shape. The field is stable in the world and volatile in the typing, so change in the data is evidence about the collection route rather than about the person.
What does most recent actually mean?
Three different timestamps commonly sit on a registration row, and teams reach for whichever is nearest.
There is when the thing was true, which for a registration is the moment the person filled the form in. There is when the row was created in the source system, which for a bulk upload of an exhibitor guest list may be months later. And there is when your pipeline loaded it, which is the one most likely to be present on every table and the one that carries no information about the world at all.
Using load time makes recency a statement about your own ingestion order. Reload a 2021 file after a migration and every value in it becomes the newest thing you have, and the survivorship rules will happily promote five year old job titles across the estate. That failure is easy to produce and hard to spot, because nothing errors.
Pick the observed timestamp, fall back to source creation, and record which one you used per field so that a later reader can tell. Where the source gives you a date and no time, note that ties become common and that a deterministic tie break has to sit underneath, which J26 covers.
The ordered rule that works
Written out, the version that survives contact with a real file has five steps.
Discard every candidate with no value, so a null can never win. Where one candidate remains, take it. Where several remain and their sources differ materially in measured correction rate, prefer the more reliable source. Otherwise take the most recent by observed time. Break any remaining tie deterministically by source rank and row identifier.
Note the ordering: recency runs fourth. On the two rows above, that produces Meixiang Chen with the 2021 phone number, the 2021 job title unless a newer non empty one exists, and the email both rows agree on. It also produces a record where every field can be traced, which is golden record construction in J27.
Changing an ordering like this on a live estate deserves a shadow run rather than a release note. Compute the new rule alongside the old one, write both results, and count the fields that move. On a file of 400,000 resolved people you might see 9,000 names change and 300 job titles, and the ratio between those two counts tells you immediately whether the change is doing what you intended. Then read a sample of 50 changed names by eye before anything is promoted, because a rule that fixes the kiosk problem may also demote a genuine married name change on a row you have not thought about. The count answers whether the change is safe at scale, and the sample answers whether it is right.
Where this stops
The correction rate that drives the source ranking is measured on corrections you noticed, which biases it towards sources whose errors are visible. Kiosk surname typos get reported because badges are printed and people read them. A wrong job title from the same source may sit unchallenged for years, so the measured exactness of a source is really its exactness on the fields somebody checks.
The deeper limit is that no rule ordering can distinguish a genuine update from a confident error when both look the same on the row. A delegate who really did change employer last month and a delegate whose employer was mistyped last month present identically, and the only thing that separates them is evidence you do not have. Where the field matters commercially, the answer is to ask rather than to infer: a pre show confirmation email that shows the current values and offers a correction link will resolve more of these in a fortnight than any rule change, and it produces labelled data as a side effect.
This week, pick your three largest registration sources and compute two numbers for each: the share of fields they populate, and the share of their rows that were later corrected. Put the six numbers in front of whoever owns the survivorship configuration. If your rule currently reads only the timestamp, those numbers are the argument for changing it, and they belong in the same conversation as conflicting values across sources in J29 and any wider unified data plan.
Questions people ask about most recent wins rule
- When is the most recent wins rule the right survivorship rule?
- For attributes that change in the world rather than in the typing. Job title, employer, phone number and postal address all move as a person's life moves, so the newest observation is usually the true one. For attributes that are stable, a newer value is more likely to be an error than an update.
- Why does a recency rule produce empty fields?
- Because the newest row is often the sparsest. An on-site kiosk registration or a re-registration typed on a phone may carry three fields where an older profile carried seven, and a rule that takes the newest row's value including its blanks will overwrite good data with nulls. Skip empty candidates before applying recency.
- What should come before recency in a survivorship rule?
- Two checks. Discard candidates with no value, so a null can never win. Then compare the sources by their measured correction rate, since a kiosk queue typed by a colleague produces more surname errors than a self service web form. Apply recency only among candidates that survive both.
Related reading
- Survivorship rules that decide which value wins on a merged attendee record
- Conflicting values across sources and how to pick one without guessing
- Golden record construction after the merge decision has already been made