CREATE TABLE ... WITH SOFT DELETES
If we like to talk about in-database processing so much, why don't we just put the actual common features in the DB, so that toggling them on or off doesn't take a bunch of code changes in app, or that you'd even be able to add soft deletes to any legacy app that knows nothing of the concept, on a per table basis or whatever.
Maybe it would make more sense to specify columns that determine if the record shows up in the results by default if set vs unset. Like archived_at timestamp, is_draft bool, is_deleted bool, is_published bool, published_at timestamp, awaiting_moderation bool, etc. ...At the expense of more implicit where conditions.
Here's a plugin for Postgraphile (GraphQL engine) that lets you do that: https://github.com/graphile-contrib/pg-omit-archived -- Kinda feels like the right abstraction.
They need to be surfaced to the product owner to decide. There may very well be reasons pieces of data should not be stored. And all of this adds complexity, more things to go wrong.
If the product owner wants to start tracking every change and by who, that can completely change your database requirements.
So have that conversation properly. Then decide it's either not worth it and don't add any of these "extra" fields you "might" need, or decide it is and fully spec it out and how much additional time and effort it will be to do it as a proper feature. But don't do it as some half-built just-in-case "favor" to a future programmer who may very well have to rip it out.
On a personal project, do whatever you want. But on something professional, this stuff needs to be specced out and accounted for. This isn't a programming decision, it's a product decision.
Also not sure what you mean by additional effort? Created_at, updated_at or soft-deletes are part of most proper frameworks. In Spring all you need is an annotation, I've been using those in major projects and implementation cost is around a few seconds with so far zero seconds of maintenance effort in years of development. At least those fields are solved problems.
Being unable to even call the shot of whether a database table should have an updated_at or soft-delete sounds like a Dilbertian hellscape to me.
in other words - YAGNI !
That's a little vague given this specific example, which appears to be about maintaining some form of informative logging; though I don't think it necessarily needs to be in the form of an DB table.
- updated_at
- deleted_at (soft deletes)
- created_by etc
- permission used during CRUD
to every table is a solution weaker than having a separate audit log table.
I feel that mixing audit fields with transactional data in the same table is a violation of the separation of concerns principle.
In the proposed solution, updated_at only captures the last change only. A problem that a separate audit log table is not affected to.
Put your documentation in doc strings where the function is defined - don’t have a separate file in a separate folder for that. It might separate concerns, but no one is looking there.
Similarly if those fields aren’t nullable, someone trying to add new rows will have to fill in something for those metadata fields - and that something will now very likely be what’s needed, rather than not pushing anything to the audit table.
Obviously your app can outgrow these simple columns, but you’re getting value now.
The absolute basics is to support snapshots and event replay. This is hardly rocket science.
Do the long walk:
Make the schema fully auditable (one record per edit) and the tables normalized (it will feel weird). Then suffer with it, discover that normalization leads to performance decrease.
Then discover that pruned auditing records is a good middle ground. Just the last edit and by whom is often enough (ominous foreshadowing).
Fail miserably by discovering that a single missing auditing record can cost a lot.
Blame database engines for making you choose. Adopt an experimental database with full auditing history. Maybe do incremental backups. Maybe both, since you have grown paranoid by now.
Discover that it is not enough again. Find that no silver bullet exists for auditing.
Now you can make a conscious choice about it. Then you won't need acronyms to remember stuff!
If you never use it, that data can be dumped to s3 glacier periodically (e.g. after 90 days).
By losing the foreign key you gain flexibility in what you audit. Maybe audit the operation and not the 20 writes it causes.
So like OP said, no silver bullets exist for auditing.
It is the point where you give up modeling the audit as part of the systems tables.
The drawbacks of this choice are often related to retrieval. It depends on the engine.
I once maintained a system that kept a fully working log replicated instance delayed by 24h, ready for retrieval queries, in addition to regular disk backups (slow costy retrieval).
I am more developer than DBA, so I can probably speak more about modeling solutions than infra-centric solutions.
I'm not saying databases are blameless. It's just that experiencing the issues they have by yourself is rewarding!
There is also a walk before the long walk of databases. Store things in text files and use basic tools (cat, sed, sh...).
The event driven stuff (like Kafka) reminds me of that. I am not very familiar with it though, just played a little bit with it once or twice.
It is tempting to supernormalize everything into the relations object(id, type) and edit(time, actor_id, object_id, key, value). This is getting dangerously and excitingly close to a graph database implemented in a relational database! Implement one at your peril — what you gain in schemaless freedom you also lose in terms of having the underlying database engine no longer enforcing consistency on your behalf.
This feels like a great unresolved tension in database / backend design - or maybe I'm just not sophisticated enough to notice the solutions?
Is the solution event sourcing and using the relational database as a "read model" only? Is that where the truly sophisticated application developers are at? Is it really overkill for everybody not working in finance? Or is there just not a framework that's made it super easy yet?
Users demand flexible schemas - should we tell them no?
Aren't you describing a non-functional approach to event sourcing? I mean, if the whole point of your system is to track events that caused changes, why isn't your system built around handling events that cause changes?
I frankly hate this sort of thing whenever I see it. Software engineers have a tendency to optimize for the wrong things.
Generic relations reduce the number of tables in the database. But who cares about the number of tables in the database? Are we paying per table? Optimize for the data model actually being understandable and consistently enforced (+ bonus points for ease of querying).
So is_deleted would contain a timestamp to represent the deleted_at time for example. This means you can store more information for a small marginal cost. It helps that rails will automatically let you use it as a Boolean and will interpret a timestamp as true.
A light switch doesn't have an atomic state, it has a range of motion. The answer to the question "is the switch on?" is a boolean answer to a question whose input state is a range (e.g. is distance between contacts <= epsilon).
Original design: store a row that needs to be reported to someone, with an is_reported column that is boolean.
Problem: one day for whatever reason the ReporterService turns out to need to run two of these in parallel. Maybe it's that the reporting is the last step after ingestion in a single service and we need to ingest in parallel. Maybe it's that there are too many reports to different people and the reports themselves are parallelizable (grab 5 clients, grab unreported rows that foreign key to them, report those rows... whoops sometimes two processes choose the same client!)... Maybe it's just that these are run in Kubernetes and if the report happens when you're rolling pods then the request gets retried by both the dying pod and the new pod.
Alternative to boolean: unreported and reported records both live in the `foo` table and then a trigger puts a row for any new Foos into the `foo_unreported` table. This table can now store a lock timestamp, a locker UUID, and denormalize any columns you need (client_id) to select them. The reporter UPDATEs a bunch of rows reserving them, SELECTs whatever it has successfully reserved, reports them, then DELETEs them. It reserves rows where the lock timestamp IS NULL or is less than now minus 5 minutes, and the Reporter itself runs with a 5 minute timeout. The DB will do the barest amount of locking to make sure that two UPDATES don't conflict, there is no risk of deadlock, and the Boolean has turned into whether something exists in a set or not.
A similar trick is used in the classic Python talk “Stop Writing Classes” by @jackdied where a version of The Game of Life is optimized by saying that instead of holding a big 2D array of true/false booleans on a finite gameboard, we'll hold an infinite gameboard with a set of (x,y) pairs of living cells which will internally be backed by a hashmap.
E.g. a field called userCannotLoginWithoutOTP.
Then in code "if not userCannotLoginWithoutOTP or otpPresent then..."
Thus may seem easy until you have a few flags to combine and check.
An enum called LoginRequirements with values Password, PasswordAndOTP is one less negation and easier to read.
Not that it really matters; deleted_at times for your database records will rarely predate the existence of said database.
MyModel.nondeleted.where(<criteria>)
which generates a query with "WHERE deleted_at IS NULL"
1-1-1970 is fine.
Many of these "we are going to need it"s come from experience. For example in the context of data structures (DS), I have made many "mistakes" that I do correctly a second time. These mistakes made writing algorithms for the DS harder, or made the DS have bad performance.
Sadly, it's hard to transfer this underlying breadth of knowledge and intuition for making good tradeoffs. As such, a one-off tip like this is limited in its usefulness.
If it's not data that's essential to serving the current functionality, just add a column later. `updated_at` doesn't have to be accurate for your entire dataset; just set it to `NOW()` when you run the migration.
But for the example of the "updated_at" column, or "soft delete" functionality, you only find out you need it because the operations team suddenly discovered they needed that functionality on existing production rows because something weird happened.
public interface ITrackable { DateTime CreatedOn {get; set;} DateTime ModifiedOn {get; set;} }
Saves so much time and hassle.
Use a popular framework. Run it against your test database. Always keep backups in case something unforseen happens.
Something especially trivial like adding additional columns is a solved problem.
Even in the best case (e.g. basic column addition), the migration itself can be "noisy neighbors" for other queries. It can cause pressure on downstream systems consuming CDC (and maybe some of those run queries too, and now your load is even higher).
"HOW DARE YOU MODIFY MY DOCUMENTS WITHOUT MY..."
So really you probably just want a reference to the tip of the audit log chain.
I like the heuristics described here. However if these things aren't making it into a product spec where appropriate, then I smell some dysfunction that goes beyond what's being stored by default.
Product need (expressed as spec, design, etc) should highlight the failure cases where we would expect fields like these to be surfaced.
I'd hope that any given buisness shouldn't need someone with production database access on hand to inform as to why/when/how 'thing' was deleted. Really we'd want the user (be it 'boss' or someone else) to be able to access that information in a controlled manner.
"What information do we need when something goes wrong?". Ask it. Drill it. Ask it again.
That said, if you can't get those things, this seems a fine way to be pragmatic.
That said, the monkey paw of this would be someone reading it and deciding they should capture and save all possible user data, "just in case", which becomes a liability.
Data has its own life cycles in every area it passes through. And it's part of requirements gathering to find those cycles: the dependent systems, the teams, and the questions you need to answer. Mindlessly adding fields won't save you in every situation.
Bonus point: when you start collecting questions while designing your service, you'll discover how mature your colleagues' thinking is.
Actually erasing data is quite hard. Soft deletes doesn't add any new lies, they just move the lies to the upper layer.
(It Is Probable That While Not Immediately Required The Implementation of Storage of Data In Question May Be Simpler Now Rather Than Later)
I've gone ahead and included additional detail in the acronym in the event that the clarity is required later, as this would be difficult to retrofit into a shorter, more-established acronym.
I find the complexity to still feel awkward enough that makes me wonder if deleted_at is worth it. Maybe there are better patterns out there to make this cleaner like triggers to prevent deletion, something else?
As for the article, I couldn't agree more on having timestamps / user ids on all actions. I'd even suggest updated_by to add to the list.
Doing this with pure 'hard' deletes is not possible, unless you maintain 2 different tables, one of which would still have the soft delete explicit or implicit. You could argue the full db log would contain the data for the former requirement, but while academicly correct this does not fly in practice.
Something like a loan could live in a production environment for well over a year after closing, while an internal note may last just a month.
Deleting rows directly could mean you're breaking references. For example, say you have a product that the seller wants to delete. Well, what happens if customers have purchased that product? You still want it in the database, and you still want to fulfill the orders placed.
Your backend can selectively query for products, filter out deleted_at for any customer facing queries, but show all products when looking at purchase history.
There are times when deleting rows makes sense, but that's usually because you have a write-heavy table that needs clearing. Yes, soft-deletes requires being careful with WHERE statements filtering out deleted rows, but that's a feature not a bug.
You might still want to show to those customers their purchase history including what they bought 25 years ago. For example, my ISP doesn't have anymore that 10 Mb/s fiber optic product I bought im 2000, because it was superseded by 100 Mb/s products and then by 1 Gb/s ones. It's also not my ISP anymore but I use it for the SIM in my phone. That also accumulated a number of product changes along the years.
And think about the inventory of eshops with a zillion products and the archive of the pady orders. Maybe they keep the last few years, maybe everything until the db gets too large.
SQL:2011 temporal tables are worth a look.
If you have no audit log(or a bad one), like lots of apps, then you have to care a lot.
Personally, I just implement a good audit log and then I just delete with impunity. Worst case scenario, someone(maybe even me) made a mistake and I have to run undo_log_audit() with the id of the audit log entry I want to put back. Nearly zero hassle.
The upside, when something goes wrong, I can tell you who, what and when. I usually have to infer the why, or go ask a human, but it's not usually even difficult to do that.
Should this be at the application code level, or the ORM, or the database itself?
* Who
* What
* When
* Ideally Why
Also you need to take into account your regulatory environment. Sometimes it's very very important that your audit logs are write once, and read only afterwards and are stored off machine, etc. Other times it's just for internal use and you can be a little more lax about date integrity of your audit logs.
Our app is heavily database centric. We push into the DB the current unix user, the current PID of the process connecting to the DB, etc(also every user has their own login to the DB so it handles our authentication too). This means our database(Postgres) does all of the audit logging for us. There are plenty of Postgres audit logging extensions. We run 2 of them. One that is trigger based creating entries in a log_audit table(which the undo_log_audit() code uses along with most reporting use cases) and a second one that writes out to syslog(so we can move logs off machine and keep them read only). We are in a regulated industry that gets audited regularly however. Not everyone needs the same level of audit logging.
You need to figure out how you can answer the above questions given your architecture. Normally the "Why" question is hard to answer without talking with a human, but unless you have the who, what and when, it's nearly impossible to even get to the Why part of the question.
I've been working on one in Typescript (with eventual re-writes in other langs. like Rust and Go), but it's difficult even coming up with conventions.
I'd counsel anyone considering event sourcing to use more "low power" solutions like audit logs or soft deletes (if really necessary) first if possible.
Anyone who has worked at a small company selling to large B2B SaaS can attest we get like 20 hits a day on a busy day. Most of that is done by one person in one company, who is probably also the only person from said company you've ever talked to.
From that lens, this is all overkill. It's not bad advice, it's just that it will get quoted for scenarios it doesn't apply. Which also apply to K8S, or microservices at large even, and most 'do as I say' tech blogs.
That's true for any other good advice you may have heard of.
I do. Each one is 8 bytes. At the billions of rows scale, that adds up. Disk is cheap, but not free; more importantly, memory is not cheap at all.
For example: as a company aspires to launch its product, one of the first features implemented in any system is to add a new user. But when the day comes when a customer leaves, suddenly you discover no one implemented off-boarding and cleanup of any sort.
Not a huge fan of the example of soft delete, i think hard deletes with archive tables (no foreign key enforcement) is a much much better pattern. Takes away from the main point of the article a bit, but glad the author hinted at deleted_at only being used for soft deletes.
Instead, just for the tables where you want to support soft delete, copy the data somewhere else. Make a table like `deleteds (tablename text not null, data jsonb not null default '{}')` that you can stuff a serialized copy of the rows you delete from other tables (but just the ones you think you want to support soft delete on).
The theory here is: You don't actually want soft delete, you are just being paranoid and you will never go undelete anything. If you actually do want to undelete stuff, you'll end up building a whole feature around it to expose that to the user anyway so that is when you need to actually think through building the feature. In the meantime you can sleep at night, safe in the knowledge that the data you will never go look at anyway is safe in some table that doesn't cause increased runtime cost and development complexity.
I'll show myself out.