> Wrt Andy, here are [1] somehow interesting views from (presumably) previous employees.
I am only seeing this now and I take the complaints about being "slightly racist and offensive" very seriously. I am checking with investors, former HR people, and co-founders. I was not made aware of any issues. If anything, I was overly cautious at the company.
I was openly transparent with our employees about every direction the company was pursuing up until the very end. The complaint that "He thinks he knows everything about business" makes me believe this person is just trolling because I was always the first to admit in meetings that I was not an expert in how to run a business. We had to fire people because of inappropriate behavior, but not because I had strong disagreements with how to run the company.
FWIW, as someone who had multiple friends who worked closely with Andy over 5+ years and at all stages of their career (both BS / PhD) those comments reek of someone with an axe to grind. All of the many anecdotes I have about Andy paint a picture of a great advisor and mentor. I suppose I should say "shenanigans aside", but if you can't separate his jokes from his academic side you need to develop a sense of humor.
> Wow, the reasons why Redis commands API suck in Andy's video (linked in the post) are the weakest ever.
In my example, the API on a key changes based on its value type. And the same collection can have different value types mixed together. You've recreated the worst parts of IBM IMS from the 1960s. However, the original version of IMS only changed the API when a collection's backing data structure changed. Redis can change it on every key!
We didn't get into the semantics of Redis' MULTI...EXEC, which the documentation mischaracterizes as "Transactions". I'm happy that at least you didn't use BEGIN...COMMIT.
You totally miss that Redis is more like a remote interpreter with a DSL that manipulates data structures stored at global variables (keys): you (hopefully) would never complain about languages having this semantics.
I don't think you understood how Redis collections work. The items are just strings, they can't be mixed like integers or strings together or whatever, nor collections can be nested.
The Redis commands do type checking to ensure the application is performing the right operation.
In your example, GET against a list, does not make sense because:
1. GET is the retrieve-the-key-of-string-type operation.
2. Having GET doing something like LRANGE 0 -1 would have many side effects. Getting for error a huge list and returning a huge data set without any reason, creating latency issues. Also having options for GET to provide ranges (SQL alike query languages horror story). And so forth.
So each "verb" should do a specific action in a given data type. Are you absolutely sure you were exposed enough to the Redis API, how it works, and so forth?
About MULTI/EXEC, when AOF with fsync configured correctly is used, MULTI/EXEC provide some of the transactional guarantees you think when you hear "transaction", but in general the concept refers to the fact that commands inside MULTI/EXEC have an atomic effect from the point of view of an external observer AND point-in-time RDB files (and AOF as well). MULTI / INCR a / INCR a / EXEC will always result in the observer to see either 2, 4, 6, 8, and so forth, and never 3 or 5.
Anyway, I believe you didn't put enough efforts in understanding how really Redis works. Yet you criticized it with weak arguments in front of the most precious good we have: students. This is the sole reason why I wrote my first comment, I believe this to be a wrong teaching approach.
> 1. GET is the retrieve-the-key-of-string-type operation.
That's a tautological argument. The question isn't what the definition of GET is, but whether the design is good.
> 2. Having GET doing something like LRANGE 0 -1 would have many side effects. Getting for error a huge list and returning a huge data set without any reason, creating latency issues.
If this really were the reason, you'd have separate operations for tiny strings and huge strings. After all, by analogy having GET return a huge string "without any reason" would create latency issues.
The examples I made are just a subset of the protection that this provides. Similarly you can't LRANGE a set type, and so forth. So this in general makes certain errors evident ASAP (command mismatch with the key type).
This does not meant that Redis would not work having generic LEN, INSERT, RANGE commands. But such commands would end also having type-specific options, that I have the feeling is not very clean. Anyway these are design tastes, but I don't think they dramatically change what Redis is or isn't. The interesting part is the data model, the idea of commands operating on abstract data structures, the memory-disk duality, and so forth. If one wants to analyze Redis, and understand merits and issues, a serious analysis should hardly focus on these kind of small choices.
Eh. What people are really arguing about here is redis’s type system. Redis’s approach has some pros and some cons. I think dismissing redis’s approach out of hand for its choices is too simple a treatment.
Most sql databases (like Postgres) require all types to be declared once, and then they do type checking on mutation. In that sense, sql is like a static language like C. But weirdly, the results returned from a sql query are always dynamically typed values, expressed in a table. Applications reading data from sql will still typically need to know what kind of data they expect back - but they usually do that type checking at runtime.
Redis flips both of those choices. It’s dynamically typed - so it won’t check your mutations. But also, you don’t need schema migrations and all the complexity they bring. And rather than having a single “table” type, redis queries can return scalar values, lists or maps. What kind of return value you get back depends on the query function. (Eg GET vs LRANGE).
If you think of a database as the foundation underneath your house, static typing & type checking is a wonderful way to make that foundation more stable. There’s a reason Postgres is the default, after all. But redis isn’t best used like that. Instead, it’s a Swiss Army knife which is best used in small, specific situations in which a real database would be complex overkill. Stuff like caching, event processing, data processing, task queues, configuration, and on and on. Places where you want some of the advantages of a database (fast, concurrent network-accessible storage) but you don’t want to stress about tables and schema migrations.
If you really hate redis, maybe say the same thing I say about Java when I teach it to my students. “I hate this, and I’ll tell you why. But there are smart people out there who disagree with me.”
If you ask me, I wish sql looked more like redis in some ways. I think it’s quite awkward that every sql query returns exactly one “table”. I’d much rather if queries could return scalar values or even multiple tables, depending on your query.
Minor nit. Some SQL databases allow you to return multiple tables. IIRC, SQL Server stored procedures can do that. Agreed its not a language feature of SQL.
That's just because SQL clients present their results that way, AFAICS. If you use a sub-query like in, say,
select * from orders where custno = (select custno from customers where name = 'John Doe');
you'll get the same result as if you'd put that scalar in your query, like
select * from orders where custno = 123456; -- John Doe's customer number
Or maybe you're right, that to SQL databases scalar values are single-row single-column tables. But so what? In mathematics, isn't any number also the single-member set of numbers that contains only that number? Where's the harm in that? (And, hey, RDBMSes are founded on set theory...)
So I don't really see what the big problem is either way. Hoping I'm not being stupid AF, maybe you could explain further?
Mathematically, they're equivalently expressive. But returning everything as a table has bad ergonomics.
Imagine the programming language equivalent. We could make a programming language where every function call returns a table. If you expect 1 return value from your function, the caller grabs the first row out of the return array, and the first column out of that row. It would absolutely work, and that its mathematically equivalent in some sense. But it would be confusing, computationally inefficient and error prone. What happens if there's more than 1 row in the table? Or more than 1 column? What happens if the type of the columns doesn't match match what you expect? What happens if the table is empty? Or you want a function which returns a two lists instead of one? We could write that programming language. But it would be pretty weird and frustrating to use.
This is the situation today with SQL. Every query returns a dynamically typed table. Its up to the caller to parse that table.
With redis, the caller expresses to the database what kind of value they expect in the query function name. (At least, list or scalar). The database guarantees that a GET request always returns a scalar value, and LRANGE always returns a list. I think this has better ergonomics because the types are more explicit.
> You totally miss that Redis is more like a remote interpreter with a DSL that manipulates data structures stored at global variables (keys):
I think he makes the point that these "global variables" are dynamically typed; you can have "listX" and then write a non-list into that same name; statically typed systems would not allow this. He makes the fairly non-controversial point that a statically typed system (SQL, other than that of SQLite) adds a level of type safety that can guard against software bugs.
> you can have "listX" and then write a non-list into that same name; statically typed systems would not allow this
Well, that depends. In most SQL databases there are many cases where supplying the wrong type of value will implicitly convert to the expected type, often in unexpected ways that can result in subtle bugs.
PostgreSQL is very very good about really never doing this, and also a scalar vs. list is pretty much a PostgreSQL case since most other relational DBs dont have a native ARRAY type. I think you're mostly thinking of MySQL that has some int/string coercion cases which are to be clear bad, but not as egregious as "any arbitrary type goes right in with no checking whatsoever.
as mentioned, SQLite breaks all these rules and I think SQLite is very wrong on this.
This is an interesting (and correct) perspective. Global variables scare us in software but we are ok with it when it comes to application state stored in a db.
There have been proposals to have global state in programming languages which function like databases with the advantages of monitoring/persistence/naming etc, but also retain the modularity of local state.
Global variables can definitely be overused, but in the right situations, they’re generally fine. After all, the filesystem is a big global variable too. So is any database. But people don’t complain too much about that.
The strongest argument against global variables is that they don’t show up in the parameter lists of functions. In that way, they’re sort of “spooky action at a distance”. And they encourage functions to be impure. But if this bothers you, you can always pass your database connection as an explicit parameter to any function which interacts with it.
> We didn't get into the semantics of Redis' MULTI...EXEC, which the documentation mischaracterizes as "Transactions". I'm happy that at least you didn't use BEGIN...COMMIT.
Hmmm, this is a subtler issue than you make it out to be, I think, though I generally agree with you. The quality issues with Redis's technical design here interrelate substantially with user expectations/perceptions/squishier stuff.
The term "transaction" is anchored in most users' minds to a typical RDBMS transactional model assuming: a) some amount of state capture (e.g. snapshot) at the beginning of the transaction and b) "atomicity" of commit being interpreted as "all requested changes are atomically performed (or none are)" rather than "all requested changes are atomically attempted".
Redis has issues with both of those, so I'm sympathetic to your statement that what they call "transactions" is mis-characterized and would be better described as "best-effort command batching".
It's poor naming/branding to call it "transactions", and I don't think it had to be this way: MULTI/EXEC "transactions" should have been deprecated long ago--in favor of Redis scripts and other changes that should have been made in the Redis engine.
First, a defense of scripts: Redis scripts are, to a certain variety of user who wants transaction-esque functionality, not ideal. Those users may be reluctant to engage with a full procedural programming language rather than the database's query language. However, there's substantial overlap between those users and the ones who will be extremely confused by and unhappy with the existing MULTI/EXEC model--they're the folks with the most specific (wrong, in Redis) assumptions of how transactions should work, and suffer the most from them not working that way. Lua scripts, unfamiliar or not, are likely less troublesome in the long run for this cohort. Specifically, requiring users to be explicit about failure behavior of specific commands via call() vs. pcall() would remove one of the worst sharp edges of the MULTI/EXEC world.
Scripts can't answer other transaction-related needs, though. Ideally, I would have preferred that Redis go in the direction of a uniform set of conditions that can be applied to most write commands. There already are conditions in Redis, but they're not uniformly available: SET + NX/XX conditions single-key writes; WATCH semantically/implicitly conditions later EXEC commands with "if version of $key matches the version retrieved by the WATCH statement", etc. If that type of functionality were made explicit and uniformly available to all or most write operations, a further chunk of transaction-related needs could be addressed. When making single commands conditional isn't enough, scripts used to atomically batch-attempt commands could be invoked with parameters used to conditionalize those scripts' internal commands, and so on.
A final simple affordance in support of transaction-ish behavior would be a connection-scoped value type: either a modifier for arbitrary commands to have them operate on an empty database scoped to the connection, or a simple list-like value for connections to "stash" arbitrary data. This wouldn't fundamentally change any semantics, but would, at the cost of some indirection, marginally reduce the need for clients complexity when buffering conditions/commands for later flush via a pseudo-"commit"-script. This is somewhat hair-splitting, though: MULTI/EXEC is already such a connection-scoped buffer, just one that stages commands and not data. My hunch is that a data-only buffer to be consumed by scripts instead of "EXEC" would be an improvement here, but I may well be wrong.
Now, the system that results from these changes is still not as ergonomic/low friction as traditional transactions, and is especially unergonomic when users have to manually capture undo state and decide on rollback semantics during the failure of script execution. As Antirez mentioned in an adjacent comment, AOF can help ensure appropriate conistency in the face of database crashes during script execution, but database level reconciliation--aka "what is the equivalent of 'rollback' for a given script"--is still on the user to work out.
But that's what we're really talking about here, isn't it? That lack of undo (that is: the ability to capture and discard transactional state a la MVCC) is at the root of most of the weird and not-quite-transactional capabilities of Redis in this era.
Antirez is totally right that adding those capabilities would have substantially complicated the Redis engine, and I believe him when he says that made it not worth it to do so. Given that, I'd have vastly preferred a Redis which embraced providing tools that work in spite of/with full acknowledgement of that lack, rather than concealing it/confusing users by mis-branding MULTI/EXEC as "transactions".
Instead of emailing me to ask if you can translate my end-of-year blog article to Chinese like you did last year, this year you decided to just copy my format? I will be sure to mention that when I release mine next week.
If all table pages exist in memory and you are using cooperative GC, then O2N can be preferable. As workers scan version chains, they can clean up dead tuples without taking additional locks.
> EloqKV brings significant innovations to database design
What is the novel part? I read your "Introduction to Data Substrate" blog article and the architecture you are describing sounds like NuoDB from the early 2010s. The only difference is that NuoDB scales out the in-memory cache by adding more of what they call "Transaction Engine" nodes whereas you are scaling up the "TxMap" node?
See also Viktor Leis' CIDR 2023 paper with the Great Phil Bernstein:
Thank you, Andy! This is Jeff, CEO and Chief Architect of EloqData. It's a great honor for us to have THE Andy Pavlo join the discussion on our first HN submission.
If I remember correctly, NuoDB uses a shared cache with a cache coherence protocol, whereas EloqKV uses a shared nothing (partitioned) cache. The former is a local read but needs to broadcast each write to all nodes. The latter has no broadcast for writes but may be a remote read. The tradeoff is evident and we are actively exploring opportunities to strike a balance, e.g., for frequently-read, rarely-write data items, use the shared cache mode.
We appreciate you pointing us to the CIDR paper. I had the pleasure of working with Phil for some time and fondly remember many discussions with Phil on various topics many years ago. To address your question, yes, we've been trying to solve the research challenges presented in the CIDR paper. The devil is in the details. We've developed numerous new algorithms and invested significant engineering effort into the design and implementation of our products. The benefits are as follows:
- Optimality: We believe we have an overall design that optimizes synchronous disk writes and network round-trips. For instance, when the design is reduced to a single node, its performance matches or exceeds that of single-node servers. As you might expect, a lot of innovation has gone into making distributed transactions as efficient as non-distributed ones, comparable to those in MySQL or PostgreSQL.
- Modularity: Our architecture allows us to easily replace the Parser/Compute layer and Storage/Persistence layer with the best existing solutions. This means we can create new databases by leveraging existing parsers and compute engines from current database implementations to achieve API-compatibility, as well as leveraging existing high-performance KV stores for the persistence layer. This allows us to avoid reinventing the wheels and to take advantage of decades of innovations in the database community.
- Scalability: The entire system operates without a single synchronization point—not even a global sequencer. We drew many inspirations from the Hekaton and your TicToc paper. All four types of resources (CPU, Memory, Storage, Logging) can be scaled independently, as we mentioned earlier. More importantly, they can scale dynamically to accommodate workload changes without service disruptions.
We look forward to sharing more technical details as we move out of stealth mode. I hope to continue this conversation with you in person in the near future.
This is from me. I didn't realize the connection to Lutris + Enhydra. It should be listed as a "Acquired Company" + "Abandoned Project". Wikipedia also says that it lasted until 2001. Usage is different from development/maintenance. I will update the entry for the old InstantDB and add an entry for this new InstantDB.
I think given that the original InstantDB died over two decades okay and is not widely known/remembered, reusing the name is fine.
https://15445.courses.cs.cmu.edu/fall2024/schedule.html#sep-...
reply