← All news and guides
Data protection

Automating DPDP Erasure: Deletion Jobs, Backups and Processor Propagation

How to automate DPDP erasure: deletion jobs that reach every system, backups that expire, processors that confirm, and the retention rules that break jobs.

Mono Team ·

A row of server racks in a dim data hall with a single open cabinet door, small status lights glowing softly in the shallow depth of field.

Short answer: automate DPDP erasure as a queue of small, repeatable deletion tasks, one per system and one per processor, all driven by your data map. A trigger (an erasure request, a consent withdrawal or an inactivity timer) creates the tasks. Each task deletes or anonymises the person's data, checks that it worked and writes a proof event. Backups are handled by letting them expire on a documented schedule and replaying an erasure list on every restore. Processors get an instruction and must confirm. Where a law makes you keep some data, the job skips it and purges it later on a date you set. The Rules do not spell out backup erasure, so write down your backup expiry. The core duties start on 13 May 2027.

This post is about the machinery: the jobs, the backups and the hand-off to processors. For the intake side (identity checks, request states, response letters) read our rights-request workflow. It is engineering guidance, not legal advice. Which law lets you keep which data is a question for counsel.

What triggers erasure under the DPDP Act and Rules?

Three different events should start a deletion job. They look alike on the surface, but they have different clocks and different evidence.

TriggerSourceWhat the job needs
The person asks for erasureSource: Section 12(3): erase unless retention is needed for the specified purpose or by lawWhat the job needs: A request ID, a verified person key, and a response to send once every task is done
Consent is withdrawn, or the purpose is no longer servedSource: Section 8(7)(a): erase on withdrawal, or as soon as it is reasonable to assume the purpose is no longer served, whichever is earlier, unless a law requires retentionWhat the job needs: A link from each data field to its purpose and consent, so withdrawal reaches the right fields
Inactivity for the period in the Third ScheduleSource: Section 8(8) and Rule 8(1) and (2): a notice at least 48 hours before erasureWhat the job needs: A per-person timer, a notice step, and a way to cancel the job when the person logs in or makes contact

Section 8(7)(b) adds a further duty: you must cause your Data Processor to erase the personal data you gave it. So every trigger has to reach vendors as well as your own systems. Section 8(1) keeps you responsible for what your processors do on your behalf.

How do you design a deletion job that does not miss data?

A single script that runs DELETE against the main database is a common starting point, and it misses data easily. The job should be a workflow with these properties:

  • Driven by the data map. Each system that holds personal data is an entry with an owner, a person key (customer ID, email, phone) and a deletion method. A system missing from the map is a system the job will never touch. Our post on the data map shows how to build it.
  • One task per system. The trigger writes one task row per entry, plus one per processor. Tasks run independently and retry independently, so one broken connector does not block the rest.
  • Idempotent. Running a task twice must give the same result as running it once. Retries are then safe.
  • Derived stores first, source last. Delete from search indexes, caches, exports and analytics tables before the system of record. You keep the person key until every lookup has finished.
  • Method per field. Delete the row, or overwrite the identifying fields and keep the rest, where the record has to stay for fraud, accounting or referential reasons. Decide which, in advance, for every table. Anonymisation only counts if the person can no longer be identified from what is left, so test that.
  • Proof at the end. Each task finishes with a check that searches for the person key and expects no result, then writes an event.

Start with a dry-run mode that lists what each task would remove, with no personal data in the output. Review it for the first few real requests before the job deletes on its own.

Where does deleted data hide after a DELETE?

Long rows of grey metal archive shelves holding plain unlabelled cardboard boxes in a dim storeroom, one box pulled halfway out on a lower shelf, soft daylight from a high window.

A successful DELETE statement often does not remove the data from everywhere it sits. These are common places it survives, with the fix for each.

StoreWhy the data survivesWhat the job must do
PostgreSQL tablesWhy the data survives: DELETE marks row versions as dead. VACUUM removes dead row versions and frees the space for reuse, but it does not normally return it to the operating system (PostgreSQL documentation)What the job must do: Rely on autovacuum and check it is running on large tables. For tables that hold sensitive personal data, check dead-row counts after the job
Object storage with versioning (such as Amazon S3)Why the data survives: With versioning on, a plain DELETE adds a delete marker and keeps the earlier version (AWS documentation)What the job must do: Delete the version IDs, and set a lifecycle rule that expires non-current versions
Event streams (such as Kafka compacted topics)Why the data survives: Compaction keeps the latest record per key. A record with a key and a null value is a tombstone that removes earlier records for that key, and the tombstone itself is cleaned out after a time (Kafka documentation)What the job must do: Publish a tombstone for the person key, and check the topic's retention settings
Data warehousesWhy the data survives: Snowflake keeps changed data for Time Travel (1 day by default, up to 90 days on Enterprise edition and above) and then 7 days of Fail-safe. BigQuery's time travel window is 7 days by default (2 to 7 days), followed by a 7-day fail-safe (vendor documentation, November 2026)What the job must do: Know the window for each dataset. Record that deleted data lingers there for a set time, and set the shortest window you can accept
Search and vector indexesWhy the data survives: A retrieval index built from tickets or documents keeps its chunks after the source row is goneWhat the job must do: Store the source record ID on every chunk so the job can delete by ID
Caches, exports and analytics copiesWhy the data survives: Copies made for reports or CSV exports sit outside the main schemaWhat the job must do: List them in the data map and give each an expiry or a delete step

AI features add copies too: prompts, conversation history, fine-tuning sets and evaluation data. Each needs a row in the task list. Store a source record ID with each of these copies too, so the job can find them.

How do you run inactivity-based erasure with a 48-hour notice?

Rule 8(1) applies to the classes of fiduciary listed in the Third Schedule: e-commerce entities with at least two crore registered users in India, online gaming intermediaries with at least fifty lakh, and social media intermediaries with at least two crore. For them, personal data must be erased after three years from the date the person last approached the fiduciary or exercised her rights, or from the commencement of the Rules, whichever is latest. Exceptions in the Schedule cover account access and stored virtual tokens that can be used to get money, goods or services. Counsel confirms whether you fall in these classes.

If you do, the job is a small state machine:

  • Eligible. A scheduled query finds people whose last-contact date, plus the period, has passed. Define "contact" once, in code: a login, a purchase, a support message or a rights request all reset the clock.
  • Notified. Send the notice at least 48 hours before the erasure date. Rule 8(2) says the notice must tell the person that the data will be erased unless she logs in or contacts you. Store the send time.
  • Cancelled or due. If the person logs in or gets in touch before the date, the job cancels. Check again immediately before running, not only at the time of notice.
  • Erased. Run the same task list as for an erasure request, and write the same proof events.

Do not build a second deletion path for inactivity. One task list means one thing to test.

What should you do about backups?

Rule 6(1)(d) asks you to have reasonable measures for continued processing if data is destroyed or access is lost, "such as by way of data-backups". The Act and Rules do not say how erasure applies to those backups. They do not say you must edit them, and they do not say you may ignore them. So the safe design is one you can explain and evidence, and counsel should confirm it.

Editing snapshots is rarely practical. A workable pattern has four parts:

  • Keep backup retention short and known. Amazon RDS automated backups can be kept for 0 to 35 days. Manual snapshots are not deleted when the instance is deleted, so give every manual snapshot an owner and an expiry date (AWS documentation).
  • Keep an erasure ledger. Record the person key of everyone erased, with the date. Keep it minimal, protect it and give it its own expiry. A hash of an email address can still identify a person, so ask counsel how to store it.
  • Replay on restore. The restore runbook has one mandatory step: re-run every erasure in the ledger newer than the backup before the restored system takes traffic. Test this in a disaster-recovery drill.
  • Say what you do. Your response to the person states that erased data may remain in backups until they expire on a stated schedule, and is not restored to use.
A close side-angle view of rows of black storage drive bays in a data centre rack, small amber activity lights glowing in a dim room, shallow focus on the nearest bays, no labels or readable text.

An alternative is crypto-shredding: encrypt each person's data under its own key and delete the key at erasure time. It makes backups unreadable without touching them, but it is heavy to build and runs through your key management. If you use AWS KMS, remember that key deletion is scheduled with a waiting period of 7 to 30 days (AWS documentation). It suits a few high-risk datasets better than a whole estate.

How do you pass erasure on to processors?

Section 8(7)(b) makes you cause the processor to erase, and Section 8(2) allows processors only under a valid contract. Treat each processor like another system in the task list, with three extra needs: an instruction channel, a deadline and a confirmation.

  • Contract terms. State how you send erasure instructions, how fast they act, what they confirm and what happens at their sub-processors. Add erasure of everything you gave them when the contract ends, subject to any retention a law requires.
  • Instruction by API where it exists. Many tools offer a delete-user call. Use it from the job and store the response. Where there is none, send a tracked email to a named address, and treat the thread as the task.
  • Confirmation is the task result. Store the vendor's reply as evidence. No reply by the due date escalates to a named owner and keeps the request open.
  • Send the retention date with the instruction. See the next section: a processor may have to hold data longer too.

If you are the processor, the flow runs the other way: you receive the instruction, act and confirm.

How do retention periods break deletion jobs?

This is where automated erasure fails in practice. The job deletes, and a rule that says "keep" is broken. Or the job refuses to delete anything and the person's data lives for ever. Four conflicts matter.

ConflictWhat breaksJob design
Rule 8(3): keep personal data, traffic data and processing logs for at least one year from the date of processing, then erase them unless another law requires longerWhat breaks: A job that deletes at once breaks the rule. The rule's own illustration says an e-book platform must keep order details, personal data and logs for at least one year, even if the user deletes her accountJob design: Take the data out of live use now. Hold the Rule 8(3) set apart. Schedule the purge for the later of the end of the year and any other required date
A sector law with a longer period (the Act's own example is a bank that must keep identity records for ten years after closing an account)What breaks: The row is both "must erase" and "must keep", so the job errors or silently skips itJob design: Store a hold on the record with the law relied on and a purge-after date. Erase the other fields
Rule 6(1)(e): keep the logs that show access to personal data, and the data, for one year unless another law requires otherwiseWhat breaks: Logs in write-once storage cannot be edited when a person is erasedJob design: Do not rewrite logs. Keep them for the period and let them expire. Our log retention post covers the storage settings
Foreign keys and shared records (orders, invoices, audit rows that point to a person)What breaks: A hard delete fails on a constraint, or removes records the business has to keepJob design: Overwrite the identifying fields and keep the record, or split the person's data into its own table

Two rules keep this manageable. First, every record gets a purge-after date computed from all holds, and the job uses the latest. Second, the job never decides legal questions: it reads a hold record that a person with authority has entered. For how holds and legal restrictions are recorded and explained to the person, see the rights-request workflow.

What evidence should each deletion leave?

If the Board or a customer asks how you handled an erasure, you answer from the record. Write an append-only event for each step:

  • request or trigger ID, and the trigger type
  • system or processor, and the method used (delete, overwrite, tombstone, key destroyed)
  • the count of records affected
  • the result of the follow-up search for the person key
  • any hold applied, with the law relied on and the purge-after date
  • for processors, the instruction sent and the confirmation received

Store references, not personal data, in these events. Run a monthly test with a synthetic person: create a profile across your systems, trigger erasure and check that every task passes and nothing is left. A failed test is cheaper to find yourself. For the wider evidence a buyer may ask a vendor for, and why there is no DPDP certificate to show instead, see is there a DPDP compliance certificate?

Erasure automation checklist

  • Every system and processor that holds personal data is in the data map, with a person key and a deletion method.
  • Triggers for request, withdrawal and inactivity all feed one task list.
  • Each task is idempotent, retries on failure and ends with a check.
  • Stores that keep deleted data (versioned storage, streams, warehouses, search indexes) have a documented fix and window.
  • Inactivity erasure, if it applies to you, sends its notice at least 48 hours before and cancels on contact.
  • Backups have a known expiry, an owner and a restore step that replays the erasure ledger.
  • Processor instructions carry the retention date, and confirmations are stored.
  • Every record has a purge-after date, and holds are entered by a named person.
  • Counsel has reviewed your backup position and the retention periods you rely on.

Mono's DPDP implementation work builds this kind of pipeline with your engineers. Mono is not a law firm, and the legal reading of each retention period goes to your counsel.

Sources: Digital Personal Data Protection Act, 2023, Sections 8(1), 8(2), 8(7), 8(8), 12(3). Digital Personal Data Protection Rules, 2025, Rule 6(1)(d) and (e), Rule 8 with its illustrations, Third Schedule, notified 13 November 2025. PostgreSQL documentation, Routine Vacuuming. AWS documentation: Amazon S3 deleting object versions and lifecycle rules, Amazon RDS automated backups, AWS KMS key deletion. Apache Kafka documentation, log compaction. Snowflake documentation, Time Travel and Fail-safe. Google Cloud BigQuery documentation, time travel and fail-safe. Last reviewed 20 November 2026.

FAQ

Does DPDP require you to delete data from backups?

The Act and Rules do not spell out how erasure applies to backups. Rule 6(1)(d) does ask for reasonable continuity measures, such as backups. A practical approach is to keep backup retention short, record every erasure in a ledger, replay that ledger on every restore, and tell people that backups expire on a stated schedule. Ask counsel to confirm the position for your business.

What triggers erasure under the DPDP Act?

Three events. A Data Principal's erasure request under Section 12(3), withdrawal of consent or the purpose no longer being served under Section 8(7)(a), and, for the classes in the Third Schedule, a period of inactivity under Rule 8(1). Each one should start the same list of deletion tasks, and each is subject to retention that a law requires.

Do you have to make processors delete data too?

Yes. Section 8(7)(b) requires a Data Fiduciary to cause its Data Processor to erase personal data that was made available for processing, unless a law requires retention. Section 8(1) keeps the fiduciary responsible for processing done on its behalf. Send each processor an instruction, set a deadline and store the confirmation you receive.

How much notice must you give before inactivity erasure?

Rule 8(2) says at least 48 hours before the erasure date, you must tell the Data Principal that her data will be erased unless she logs in or contacts you. It applies to the classes in the Third Schedule: large e-commerce entities, online gaming intermediaries and social media intermediaries. The job should cancel if the person makes contact first.

What if a law says you must keep data that a person asks you to erase?

Sections 8(7) and 12(3) both carve out data that a law requires you to keep. Erase everything else, take the held data out of live use, record the law relied on, and schedule a purge for when the hold ends. Rule 8(3), for example, sets a one-year minimum for personal data, traffic data and logs of processing. Counsel decides how far each rule reaches.

Build erasure that works across every system and vendor.