We saved $50k/year with a Go microservice coded in a hackathon
movio.co
movio.co
What we have here is a classic dbms problem and no one at Movio seems to know how to deal with that. Instead of migrating from Mysql to something serious (Postgres) they move to some columnar DB no one has heard of. Nevermind that postgres and a reasonably priced DBA and a little thought put into their data model/queries could probably handle all their issues.
Sorry for the snark, cheers on a successful product.
As has PHP. "popularity" isn't really evidence for it being a "serious" tool, is it?
The questions I usually ask myself when evaluating database solutions is:
* Does it accept invalid data?
* Does it change data on error?
* Does the query planner change drastically between minor versions?
* How strong is transaction isolation? can I create constraints, columns or tables in a transaction?
* Does it scale vertically above 40~ CPU threads and 1M IOPS?
The answer to all these questions, for MySQL is "No". You could argue the value of some of them, but a lot of them highlight architectural or development procedural misgivings.
Most companies are not alive "because they chose mysql over something else" they're alive because they have "good enough" tech to get the job done. The job that they're trying to accomplish is the thing that makes them successful.
Uber isn't super huge because it used a specific database technology. It's huge because it's good at marketing, it's providing some value to people.
If it silently corrupts data over years and gets discovered much later. (As was the case with my previous company, an e-commerce retailer that lost large chunks of order history)
> How do you know that similar or worse problems did not exist in alternative solutions at the time it was implemented?
Because I've been working on database solutions for over 10 years, there are problems in other software but I consider data loss to be worse than any of them. For example the autovacuum in postgresql 8.3 and before was mostly garbage which ended up bloating highly transactional databases. But deleting data when you fail a constraint is worse.
Still, if you are making choices for yourself, you don't choose mediocrity and hope to muddle through: you choose excellence. Choosing MySQL isn't choosing excellence.
Do you have objective or quantifiable data and references upon which your opinion is based, _and_ is universally applicable to any arbitrary problem that a SQL database might be an appropriate solution for?
I do not think MySQL is a technological debt as in 80% startups moving to the different solution is cheap and non-problematic. The LAMP is good enough and quickest/cheapest for the majority of tech companies.
You can solve issues in your application if you know there will be issues like these, knowing the pitfalls and drawbacks of a technology is certainly noble- but if you do then why not choose something that follows principle of least surprise. (There might be reasons).
I would never claim that you should move everything from MySQL if you use it. However if you care about data consistency ensure that you change the defaults, engage strict mode, ensure that your application has no bugs in handling data.
This is actually hard to do correctly, it's overhead in development that you shouldn't be caring about. Just choose something that has sane error conditions and the problem vanishes.
To start with, this is still true today: https://vimeo.com/43536445 Despite being 6 years old, strict mode is still required.
Anything prior to MySQL 5.7 will accept "0000-00-00 00:00:00" as a valid date, 5.7 will not (which is sane) however this means migrating from 5.6 -> 5.7 just got a little harder.
In fact it wouldn't validate /any/ date so it would assume every year was a leap year and febuary always had 29 days.
Regarding the query planner: https://bugs.mysql.com/bug.php?id=74602 This affected my prod mysql/zabbix installation https://support.zabbix.com/browse/ZBX-10652
Regarding performance: This is what I found from my own experience: I was given the task of testing the limits of MySQL, MySQL was the chosen technology and I was no involved in making that decision so- whatever. We were given 10 servers, with 40 cores (2014-2015~) 128G of DDR3/ECC and 8 SATA SSDs in RAID-0 with 1G of RAID cache for write-back. We managed to get MySQL to bottleneck pretty quickly, our queries involved a lot of binary data so we should have been raw IOPS bound, but we weren't we were memory bound. So we replaced the memory allocator with a faster one (jemalloc) and we get a 30% performance improvement. We suspected that the kernel sockets implementation was slowing us down so we compiled a custom "fastsockets" linux kernel. The improvement was around 4%, but we were bottlenecked on memory. After doing a full trace of what MySQL was doing we saw that InnoDB was spinning on a lock quite a lot. I asked if we could try other SQL solutions (MSSQL/PostgreSQL) Postgresql was first chosen because we could just install it, no license and no OS change... it was twice as fast as the optimised MySQL installation out of the box with a stock CentOS6 kernel.
We never even bothered testing MSSQL because PostgreSQL met our performance targets, we were now IOPS bound.
-- More anecdatum:
Regarding data consistency we (tried) to migrate to postgresql for performance reasons in 2014 (my previous company), and failed because MySQL had been corrupting our data very slowly and silently for many years (corrupting meaning not honouring NOT NULL, not honouring type safety, allowing invalid dates, inserting data on error) So far in that actually reimporting the output of `mysqldump` would not work.
Isn't it? I thought that today()-(2018 years, 4 months and 10 days) would be approximately that date? Maybe you prefer +0000 vs just 0000?
'ISO 8601 prescribes, as a minimum, a four-digit year [YYYY] to avoid the year 2000 problem. It therefore represents years from 0000 to 9999, year 0000 being equal to 1 BC and all others AD. However, years prior to 1583 are not automatically allowed by the standard. Instead "values in the range [0000] through [1582] shall only be used by mutual agreement of the partners in information interchange."
To represent years before 0000 or after 9999, the standard also permits the expansion of the year representation but only by prior agreement between the sender and the receiver.[19] An expanded year representation [±YYYYY] must have an agreed-upon number of extra year digits beyond the four-digit minimum, and it must be prefixed with a + or − sign[20] instead of the more common AD/BC (or CE/BCE) notation; by convention 1 BC is labelled +0000, 2 BC is labeled −0001, and so on.'
https://en.m.wikipedia.org/wiki/ISO_8601
Now, if mysql accept, but can't store such a date, I understand that it's a problem.
Nontransactional DDL alone is sufficient to classify it as a toy DB for me. Yes, I've been personally bitten by it.
[0] - https://dev.mysql.com/doc/refman/5.7/en/cannot-roll-back.htm...
And https://mysqlserverteam.com/new-defaults-in-mysql-8-0/
Previously on HN: https://news.ycombinator.com/item?id=5122299
Personally, I only use MySQL and derivatives where I have to (basically WordPress.)
[1] https://stackoverflow.com/questions/4711447/oracle-ddl-and-t...
... have dedicated hundreds of engineers and millions of dollars to nothing more than keeping MySQL up, running, and not crapping the bed every time someone looks at it funny. If you can afford that resource expenditure, by all means go nuts with MySQL. Most companies can't and would be far better served by something which doesn't need that amount of handholding to serve its basic purpose.
I think by the time Postgres sorted itself out into a more user/admin-friendly system (which is still fairly recently, really), MySQL had pretty much conquered the "quick and easy" mindshare and was deeply embedded almost everywhere.
And if you've spent millions of dollars architecting your systems such that MySQL's flaws aren't killer issues, there's very little financial benefit to switching, I guess.
Following on from that, I suspect a lot of large companies use MySQL because they always have, not because it's actually any good. For example, Basecamp used MySQL while I was there, but I never met a single Sysadmin there who would use it over Postgres if they were to start a new project.
MySql and PHP are good. They do the job they were designed for in a cost effective way and of course that means there will be trade offs.
Sure, you can knock up a contact form in it really quickly, but that ease of use hides significant dangers.
I've programmed it, it's a terrible language.
Eventually Postgres caught up in most things, and the delay was in some part because of implementing those features "correctly" and with more thought, but it's still a delay that hurt the uptake in the early days.
Most E-Shops will be fine running MySQL or MariaDB.
The one thing PG excels at however is that you can tune it much more to your workload and it allows tuning the workload much more finely than MySQL/MariaDB. That and the ability to extend PG arbitrarily (try adding native functions to mysql without recompiling) via the C-FFI offered. You can write and define your own index methods that let you use an index that is perfect for the workload or you can add a new data type to support a new input with validation.
You can sink a lot of work into getting the most out of a PG database, MySQL not so much. But again, for most people MySQL will provide the same (or even better) performance than PG. (I still trust PG over MySQL after MySQL nulled out all entries of a table with only NOTNULL columns after a nasty crash)
Like, how often one needs to rollback a DDL statement? I did that like... never.
And, what is the use case? Like, you added a column to a table by accident? Well, that will not break anything, so no harm done.
That is way different from regular dml rollback which may recover 1bn records and save your life :)
For example, you had a "color" column on a table, for a new feature youre now adding the ability to have multiple colors. You're going to create a new column, create a new table, populate that table, and drop the old column. If anything fails during that process you'd like to be able to roll back.
There is a concept or "forward-compatible change". Basically, you don't do things that will break your software.
Example, you don't add a NOT-NULL column unless you can give it a good DEFAULT value, to make it work.
Also dont' drop columns until the software is ready for it, etc.
If you have a decent ORM, it will compare your "how it needs to be" sql schema with the "how it is" schema. Then it will generate appropriate "ALTER TABLE ...." "CREATE INDEX " etc statements. Note that this is automated and you never need to type SQL statements to achieve that.
All together in the last XXX years, I did not really need to do a rollback on a dml statement.
And that criteria of technologies could be summed into one sentence. Use something boring. No Hyped programming languages / DB / tools allowed.
Of course some would argue you would be doing it wrong even if it was using old tech / programming / tools. Well yes, but you have a sea of recourse and expertise there to ask for help. Instead of spending energy and time doing figuring it out.
Of course if your company is all about tech innovation, AI or something cutting edge there surely you will have to tried something new. But 80% of those startup aren't.
By choosing at most one new thing, you can better control for how your stack should work and how you expect it to respond to certain unexpected circumstances, which means you should be able to more effectively solve issues as they crop up than you'd be able to if you were using multiple new technologies.
> Embrace Boredom.
> Let's say every company gets about three innovation tokens. You can spend these however you want, but the supply is fixed for a long while. You might get a few more after you achieve a certain level of stability and maturity, but the general tendency is to overestimate the contents of your wallet. Clearly this model is approximate, but I think it helps.
> If you choose to write your website in NodeJS, you just spent one of your innovation tokens. If you choose to use MongoDB, you just spent one of your innovation tokens. If you choose to use service discovery tech that's existed for a year or less, you just spent one of your innovation tokens. If you choose to write your own database, oh god, you're in trouble.
> Any of those choices might be sensible if you're a javascript consultancy, or a database company. But you're probably not. You're probably working for a company that is at least ostensibly rethinking global commerce or reinventing payments on the web or pursuing some other suitably epic mission. In that context, devoting any of your limited attention to innovating ssh is an excellent way to fail. Or at best, delay success.
But then ... you have to find, attract and hire good developers. That's already difficult, adding an extra layer of 'boring technology' will make this task even more challenging.
The only thing I'd fully agree on is that using Go saved us a lot of resources as well. It's an awesome choice for stuff like this that needs to be reasonably performant as well as being simple, understandable and reasonably fast built.
The fact this had to happen in a hackathon suggests a typical disconnect between management and development (and probably poor prioritization by management). Because development knew this was a problem and how to fix it (evidenced by the fact they fixed it), but it took removing management (aka a hackathon) to give development the space to fix it. And now the company pats itself on the back for having the vision to host a hackathon instead of structuring and prioritizing correctly in the first place so this would just get fixed on the clock.
I do think the author's takeaway about the value of simplicity and pragmatism are on point, but that applies not just to code but to management as well.
This is just trolling
I don't mean it to be conformist, but it's easy to forget some things are actually hard when you are very clever or old enough to forget how it was like when you were still learning too.
The key, the whole key, and nothing but the key so help me Codd.
Also said as... "In Codd we trust."
If none of these DB jokes mean anything to you, take a DB concepts class at a CS university. There's a lot of great research going back 50 years and you can learn a great deal about why things are the way they are (tuple algebra and calculus). And before changing anything for something you think may be better, you should fully understand what you are giving up.
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.86....
Not that I disagree, but to be fair, I've seen plenty of tech ignorance with experienced and older engineers as well that has been pretty crippling.
It often goes like this: Oh snap, we encountered a problem! Lets find a tool, framework, language that promises to solve a similar sounding problem. Now we have a problem with a layer of abstraction on top. Soon to be two problems. Lets find a tool, framework, language to solve both of them ...
It is a spaghetti to the wall approach, where you just throw a bunch of things at your problem hoping that something sticks. And who cares how long it will stick.
Secondly as a developer I think in start-ups dedicated db experts are way underrated. Sure your fullstack devs can cobble together some tables, changing them 15 times a day to accommodate business requests and slap indexes on everything that gets slow. That is also the way to get into trouble once you scale, and instead of reflecting why this is, people reach for the bowl of pasta.
I was no different, when just starting out. I thought my biggest strength was, how quickly I can come up with easy "solutions" for any problem the company had. Took me years to realize how silly of an approach this is.
e.g. I remember using a dedicated jenkins environment to run continuous, scheduled integration tests for my service. When the architect found out, he immediately sent me links to software packages that are dedicated to running continuous tests. I asked whether he had any experience running these new packages and if he would be willing to set it up/maintain it.... radio silence.
>It often goes like this: Oh snap, we encountered a problem! Lets find a tool, framework, language that promises to solve a similar sounding problem. Now we have a problem with a layer of abstraction on top. Soon to be two problems. Lets find a tool, framework, language to solve both of them ...
is absolutely real, I've actually seen in happen both in projects I was in, and heard or read about.
This Rich Hickey video is somewhat relevant, IMO:
Tech Video: Rich Hickey: Hammock-Driven Development:
https://jugad2.blogspot.in/2016/03/tech-video-rich-hickey-ha...
You have to watch it at least part way through to get some of the better points in it, although the whole thing is good.
The difficulty I find is, identifying the moment to leave the hammock again in a startup enviroment. To what degree do you need to understand a problem before you take action. If you try to understand it 100%, you'll never get anything out there.
But I'm already very happy that I was able to convince the business side of the company of the approach in a brief talk about it and they now referrer to "the hammock" themselves :)
Agreed. The problem, though, (and I'm painting with a broad brush here) is that the erring tends to be much more on the side of not trying to understand much or at all, of the problem, before jumping into action. I think a lot of it is due to peer pressure and wanting to be "seen" by peers and bosses (and VCs) to be doing stuff, as opposed to really getting things done better in the medium term, even if in the short term it looks like you are not acting but "only" thinking or analyzing or designing stuff. Hence my comment in that post I linked to, about "we have to ship next week". All too common - been there, seen a good amount of that. In fact, this subthread between HN user jacquesm and me just recently, is basically about the same point, although described in different words:
https://news.ycombinator.com/item?id=16774234
>But I'm already very happy that I was able to convince the business side of the company of the approach in a brief talk about it and they now referrer to "the hammock" themselves :)
Cool :)
Some time ago, I thought it was <easy> to write code to do things. By now, I mostly ponder how I put things into postgres/kafka|rabbitmq|../memcache|redis|.../elasticsearch/neo4j so I can reduce everything to good queries into these systems.
I have 15 years of experience and can built a decent clean system using "boring" technologies. But all the decent paid work where I live is maintaining big balls of mud with tech that was obviously peak hype when it was chosen, and nothing done according to best practices because of that would require sticking with a tech and learning it properly. Its quite frustrating.
Then we have the interview process where people expect me to give up my weekend for their coding test and can't even be bothered to give you feedback afterwards. Or some ridiculous algorithmic nonsense that has no relevance to the job. Getting bored of it all.
Select loyaltyMemberID
from table
WHERE gender = x
AND (age = y OR censor = z)
Sure, the filters can be arbitrary but with an ORM it's really really simple to build them up from your app code. The Django ORM with Q objects is particularly great at this.
Obviously I'm armchairing hard here but it smells like over engineering from this post alone. Stuff like this is bread and butter SQL.
Edit: I've just read the query in the post again and I really can't understand why you would write it like that. Am I missing something here?
Seems like a fundamental misunderstanding of SQL rather than a particularly hard problem to solve.
Ten or fifteen years ago, sure - a DBA would look at a query plan and figure out how to do it properly. Worse case you'd slap a materialized view in and query that.
But this is 2018! Programmers don't want to treat the database as anything but one big key value store ;)
SQL databases are amazing, robust examples of engineering. They are your friends and they're the appropriate choice for the vast majority of software. They are not outmoded or passe. Though I acknowledge there is a separate use case for K-V stores, I almost want to make policy preventing their use just because I know so many developers will abuse them badly and then stare back at you blankly during the semi-annual massive downtime event, muttering something like "Well, it's based on research at Google, so I'm sure there's a way to recover the data..."
If your columns have types, how can you encrypt them using custom keys for each one? Is it possible?
I want the keys to reside on the client and search for encrypted data. Basically single row lookups at a time.
I am however extremely wary of doing it that way. I don't know your requirements of course.
What I really need is just a huge table with two fields: “token”, “content”
And the token is basically the primary key but encrypted with whatever encryption.
You could even do foreign keys this way.
Hmm I suddenly have an idea. What about a layer above the database that basically enforces foreign keys and joins in this way to support end to end encryption? The content would reference ENCRYPTED foreign keys. Only clients would decrypt stuff.
Sounds more like a key value store and less like a relational database. Although you can store key value data in a relational db of course, there may be a better tool for the job.
If it’s transactional, I recommend keeping calculations only if you need to access summarized data frequently. For example, if you are tracking inventory by storing the history of transactions that occur into and out of inventory, it’s trivial to find out how much of each item you have in stock at any point in time by doing a sum of the change in quantities for each item type up to that point.
If you were usually interested in the “current” count, it would be expensive to perform this sum every time, so instead you could keep a separate table for calculating the running total of inventory per item and referring to that. Keep this table up to date through the use of triggers on insert events (Note that your log of inventory transactions would thus be an immutable stream of events)
An example of something not worth storing is derived data that is a combination of separate columns in a table. For most queries it’s probably trivial to be lazy and wait to perform such a calculation until you actually need it. If you still want to have a ready made “table” that has all the computations you want already entered as columns for easy querying, use a view. If you find yourself making liberal use of views, you’re on the right track.
If you are using a separate data warehouse for data analysis, then precompute and denormalize as much as you can.
*Disclaimer: most of what I’m talking about is from a Postgres perspective.
It's short so you actually read it and possibly reread it. It's to the point. It has pretty pictures. And it had directly applicable advice.
My early education on databases always seemed to follow a "how do we make a database do this?" rationale rather than "what data do we need to store to support these features?", which I think leads to a software design that is too strongly coupled with the database. Software modules end up dependent on database features, or table structure, and refactoring or switching data stores becomes more costly.
Instead, start with a simple in-memory data store - a list of objects with some interface for accessing them, will probably your starting point. Add some basic serialisation/deserialisation features (CSV, JSON, etc) when you get past initial testing and require some persistent data. Then, once you have your API in place and your software design is stabilising, you should be able to map that data to a database fairly easily:
* The primary structure maps to your main table
* Child structures become additional tables, with foreign keys
* Data used to lookup records can be indexed for better performance
Beyond that, you should profile/benchmark your application to find what needs to be optimised, and then investigate whether your software design or your data store should be doing the optimisation.
Let your software's features influence the design of your database. Don't let the database's features influence the design of your software.
1) It took a weekend to complete. The friction was building for far longer. There's a cost to that, esp if it effects customer satisfaction and retention. They didn't refactor for fun, did they :) How many dev teams aren't so lucky? Is this article a no choice outlier, or a best practice?
2) My comment wasn't directed at the article but on another comment that blamed the developers. These problems should be owned by ownership / leadership / management more and engineers less.
3) That said, hire a DBA? I don't think that's necessary.
If you are claiming to be full stack you better be prepared to go all the way.
In any case, having a basic understanding of the next level up (the electron) has proved quite useful to my career, otherwise I wouldn't know how turning things off then on again affects the machines I'm working with.
Granted, such knowledge isn't immediately useful since it's something I or anyone is likely to do but it grants insight into systems. I know roughly how a query optimizer does and what it can, and more importantly, can't do.
When you know a system you can optimize for it. When you don't know a system you can only follow someone else's advice on how to optimize for it.
Nitpick: That's actually either three ^w or just one ^w, depending on how your WORDCHARS is set up. :)
As it stands the SQL query is quite silly. It gets a list of every user ID that is included by each filter and compares which ones are in the filters you want and not the filters you don't want. Much better is to pass the filters into SQL, let it figure out which users match the filters you want and not the filters you don't, and just use that result.
This leads to lots of unions in advanced queries, and makes filtering harder. Some databases even calculate column block statistics to optimize these queries by doing less IO even for seeming table scans.
Why not one table with all customers and one column per property? There are a few reasons, having to do with anything from MySQL sucking at schema alters for really big tables, to expectations of Enterprise customers.
The general form of this is:
Select loyaltyMemberID from table WHERE V1_1= x_1 OR ... OR V1_n=x_n) AND (V2_1 = x_2_1 OR V2_2=x_2_2 OR ... V2_n=x_2_n) AND ... AND (Vn_1 = x_n_1 OR ... OR Vn_n= x_n_n) (some of these n's should actually be m_i's but I was lazy)
There may be some ability to optimize this in a number of ways but optimizing one example is not optimizing the general form. I can easily see how technology change could be a cleaner solution.
I totally get that, but isn't that the point of the query optimizer within the database itself? Why are you trying to outwit it? It should select the right indexes, provided the columns are indexed, and "do the right thing(tm)". It might take a bit of cajoling but they seem pretty good at this. Postgres collects statistics about the distribution of values themselves within the table to guide its choice of index, so in theory it could rewrite the boolean logic to use a specific index if it's sure that it will eliminate a higher % of the rows than another plan.
In any case, it seems the SQL they posted is a bit off. Why nest each individual filter as a UNION? If you wanted to go down the UNION route couldn't you do each individual group as a UNION, with standard WHERE filters?
Oh I've seen this happen a lot. Somewhere along the line, often from a DBA, it is decided that sql in an app is evil and that everything must be in a stored proc. Then instead of some simple string concatenation you have to jump through hoops like this.
> no chance of SQL injection then.
You know you can have sql injection attacks inside stored procedures? If you think stored procedures are a panacea then you don't understand the problem you're solving.
I'm not using something like Entity Framework and the CRUD apps I mostly wrote at work were well suited to a few simple sprocs.
The time it takes to write an ALTER script to change something pales in comparison to the two week change control process anyway...
id, name, age, gender
1213, fake, 60, female
property_id, user_id, value
1 (assume age), 1213, 60
2 (gender), 1213, female
Edit: Had some bad attempt at writing this query but it's rather late and it made no sense.
SELECT DISTINCT loyaltyMemberID
from members as m
INNER JOIN properties as p1 on m.id = p1.user_id
INNER JOIN properties as p2 on m.id = p2.user_id
INNER JOIN properties as p3 on m.id = p3.user_id
AND (p1.prop = 'gender' AND p1.value = x)
AND ((p2.prop = 'age' AND p2.value = y) OR (p3.prop = 'censor' AND p3.value = z))And they mention in the post that most queries don't use that many fields.
SELECT DISTINCT loyaltyMemberID
from members as m
INNER JOIN properties as p on m.id=p.user_id
WHERE (p.prop='name' AND p.value = value) AND
...etc.If you get records back with prop1 then you have to remove those records from results based on another record.
There are multiple ways to accomplish this but it can't be done with a single join and simple where clause.
You could have
property_id, user_id, value(string)
1 (assume age), 1213, 60
2 (gender), 1213, female property_id, user_id, value_str, value_int
1 (assume age), 1213, null, 60
2 (gender), 1213, female, nullStill, a JSON column will arguably be faster than a pure KV table since you can more efficiently query it, especially any non-JSON columns.
So values in JSONB columns can be indexed nicely, but the statistics can be much worse than for non-JSONB columns, which can lead the query planner astray.
"Relational databases don't scale"
Well they worked fine for decades before Mongo's marketing claimed so.
Normalize it properly. If this is not possible, ensure that both the userid and the property are indexed
I think you're right. Oh ... my ... god ...
I wish I could say this is the worst example of a database schema I've ever seen, but it isn't.
Technology cycle:
X gets invented -> idiots abuse it -> X "is bad" -> Y (strictly worse than X) is "so much better" -> idiots abuse it -> Y "is bad" -> ...
In Mysql, it is JSON data type [1], in Postgres JSON/JSONB [2].
Creating indexes across them is doable, through a workaround (involving what is generally referred to as "VIEWS", but can be called calculated columns or something like that).
And, frankly, in the worst case for indexing, these databases still perform comparable to key-value stores in speed (especially SQLite).
[1] https://dev.mysql.com/doc/refman/5.7/en/json.html#json-paths
[2] https://www.postgresql.org/docs/9.4/static/datatype-json.htm...
Using an EVA, you can have a "keys (id, key_name)" table, and then only use the IDs in the values table, reducing that waste.
By the way, you don't need views for indexing on JSONB fields, it's supported out of the box in PG.
And I don't think you will necessarily get better performance with json fields vs an EAV model. Yes, you can index json fields by creating virtual views, but that requires that you know the field ahead of time. With an EAV model, you can have your values table indexed and then join.
But I am excited to start using the json field types. In many cases, it will really simplify things over the traditional EAV stuff.
[0]: https://en.wikipedia.org/wiki/Entity%E2%80%93attribute%E2%80...
"can be useful in various scenarios, such as rows with many attributes that are rarely examined"
They are querying these quite heavily, they aren't just random attributes they need to retrieve.
[0]https://en.wikipedia.org/wiki/Entity%E2%80%93attribute%E2%80...
Then years later I ran into the guy who had to do some work on it after I left that company. I must have scarred him pretty badly, because he remembered it enough to bring it up as pretty much the first topic after the obligatory 'hey so what are you up to nowadays'. When I think back about it now, it was a cringey solution - then again, this was at a company where nobody had ever heard of a 'database index' (or if they did, never mentioned or implemented them).
Without knowing the rest of their stack, or what their data ingestion looks like, I think your query is oversimplified. If they are doing a union, then it's likely they aren't querying one table, but they are querying multiple tables. The article mentions that individual customers had as many as 500 million rows. Likely each customer has their own set of data they also pipe into the system. Next their own custom query language may support more complex algebra than standard equality.
IMO, the article doesn't sufficiently describe the problem for us to understand why their solution works. To you and I there are 100 other solutions they could have tried that seem simpler than the one they presented.
It's less likely that they overengineered - we are probably just underinformed.
Based on 15 years in software companies in the valley it's much less likely that this isn't over-engineered. Nearly every decision I've seen chasing technology hype has been based on ignorance of existing solutions.
https://en.wikipedia.org/wiki/Entity%E2%80%93attribute%E2%80...
Likewise you can get away with a full document GIN index.
I played around with some basic report stuff at work last year, the EAV data on my local machine, the report took ~7 seconds to run. I shoved the same data into PostgreSQL as JSONB, indexed it just as full doc cos I was lazy, the same report took ~80ms.
Obviously this isn't 'proof' my dataset was only 1.5m by 15m records. But with my limited knowledge i do believe it would perform better, I don't know how much better... but I think better...
It's quite fun.
What I am saying is, I really do not want a situation on my hands where the juniors that I work with, or most of the intermediates, and even a few of the seniors and leads, are writing raw SQL or query objects. Most of these folks have n years of experience in web and desktop application development and couldn't give you a passable answer to simple questions like, "What's a database index?" I know this isn't isolated to my current employer, or former employers, and I've seen it in other organizations where I've done some consulting on the side, and all of these folks I'm talking about here have largely worked else where in the past, too. And this in itself leads to other third-order effects, like the "SQL wizards" who get asked all of the "tough" SQL / database questions.
I want to stress that I understand the point that you're making, and I do agree with it, and of course, so do many (all?) ORM authors themselves, but I think the advice is wrong and is prone to take you to a much worse situation. I think we have an obligation as people who do grok SQL and databases to gently introduce our less experienced co-workers to the idea that ORMs are not a panacea to all database interactions, but until the companies we work have enough of an incentive to give us that sort of time and empowerment then I, for one, am going to recommend ORMs for everyone for everything unless they really, absolutely, demonstrably know what they're doing.
One of the problems, in my opinion, is that SQL isn't "cool" or hip and by many seen as not important to learn. While the new fancy Javascript based language or framework which nobody use and that will be replaced next week is much more important to learn.
Btw, get of my lawn :) /end old man rant
> could also ORM code with N+1 queries
Oh, they absolutely do, and when we're lucky they actually catch them on their own before they get to code review. Some folks reach for tools like Bullet [0] and, that's great, but unfortunately, sometimes they treat that tooling like the Holy Gospel. They develop an over-reliance on them as if those tools exist to offload critical thinking. Drives me crazy... in my experience, it's been hard to combat this type of thing, too. The pace of "agile," the calculus between paying down technical debt and mentoring and progress, I don't really know why but I haven't had a lot of long-term luck.
> One of the problems, in my opinion, is that SQL isn't "cool" or hip and by many seen as not important to learn.
I think you're really right about that. I happen to like writing SQL quite a bit and I take a little bit of pride in that I kind of sort of actually understand a little about what is going on in the database and even then I neglect that skill. I picked up copies of both "SQL Anti-Patterns" and "SQL Performance Explained" based on recommendations from this thread and am eager to get in to them this weekend. Still lots to learn... And, I have some SQL problems that I can see coming up over the horizon today and I hope this gives me the edge I need to start grappling with them sooner rather than later.
- they didn’t have the expertise to actually fix the SQL. That query smells bad. The data model smells bad. For some reason HN is always superstitiously afraid of letting developers touch the database, but if you don’t let devs touch the database enough you end up with this sort of thing; or that crap data model with properties in rows instead of columns, because oh god, we can’t let devs actually do DDL so we’d better make it all really flexible (and incredibly slow because it’s a misuse of the database). I mean, implementing your own result caching mechanism? I don’t know about MySQL but surely it has its own caching mechanism (Oracle does) that isn’t being used because the query is bad.
- project management probably had no interest in fixing the performance/incorrect data problems, and devs were expected to do it in their own time.
In a way though this makes me feel better, other people are dealing with these problems too and their overengineered solutions work and keep the company running, I guess mine will too :)
Seems like a red flag.
I'm all for companies releasing technical blog posts, but there's some really strange framing here.
This is actually a story about how decisions get made, and how better ones can be made. Reading a company's mea cupla tells you they are well-informed and well-intentioned.
This is not a story which presumes good decisions were made and "the (tiny, startup) database company went bust". That's their framing. Yikes.
How they do it? I don't know. They are using Xeon processors and not i7 like some others.
The Server Auction is basically servers they previously used and were freed up (because a customer didn't need it anymore, for example).
It's somewhat similar to what OVH is doing with KimSufi or SYS.
You can see there 64Gb boxes with NVMe drives startin around $160 USD per month. Not as much savings as with Hetzner, but still beats 1:1 price of AWS.
But if all you use is RDS, EC2 and S3 - it is quite easy to move to dedicated hardware once your service is up and running on AWS. Unless your load is not jumping around 10x all the time of course (then you will need dynamic scaling which is not quite possible with dedicated hardware)
The execution plan will tell you how expensive is each bit of your query and help you adjust it.
From there, if things are not getting better, you have a lot of alternatives:
- Consider creating an index
- If the value doesn't change often, consider writing it into another table or caching it.
- Replication, partitioning, sharding, changing the schema.
- Reconsider the requirement being implemented in order to have a more scoped query or to perform the query less often.
Then... OLAP is not OLTP. If you can, do reporting in another database.
Finally, creating your own project in the end may not save you $50,000. How about maintenance? tooling built around it? integration costs? documentation? usability? new hires having to learn about it? You can hire people that already know SQL without having to incur that cost yourself. All the tooling is built, battle-tested and readily available. Plus, skills related to internal tools are harder to trade in the market because they're harder to verify and less transferable.
Long term high intensity output will lead to burnout, even if the salary is 10x people would struggle and crash. Pushing at 100% full enthusiasm is like sprinting, it is not possible to maintain that intensity for very long. It can be fun, it can be productive, but the wiser approach has the long-term and end in mind.
Regardless of all of that, in my experience a typical impressive hackathon project is still just a barely working demo that benefitted from a significant amount of research and planning beforehand, and will require an even greater amount of hardening and polish afterwards.
There is no magic, it's just a vastly different kind of work environment with both inputs and outputs incomparable to day-to-day work.
A proposal like this could easily have been seen as the developers wanting to test out a technology that was not approved or with a good business case. That business case is usually something that only sales/product can sell. The barrier to listening to developers is higher because they are assumed to not know enough about business.
You may have an "agile" environment, but you often need a very good reason to not pick the next item from the backlog, which was not created and maybe not even prioritized by you.
In those companies, the hackathon may be the only time developers can present their ideas.
Maybe I'm crazy, but just looking at that query it seems like there's definitely room for improvement with the SQL alone. Unless the "..." is hiding something I'm missing?
The numbers they quote (5M members, 100M transactions) are tiny for any modern data warehouse. Many solutions would run these in sub-second speeds without changing the SQL at all, and it would be far better than building a quasi-SQL engine in Go.
Actually for the occasional querying + caching that they have, something like BigQuery or Snowflake data would be even cheaper with basically 0 operational effort.
"NYSE moves from Solaris to RHEL and gains a 800% performance benefit".
While I don't doubt a brand new RHEL has more performance optimizations than what is actually a SunOS 5.2, the guys benchmarking should have also said that the original hardware was the equivalent of a PIII and now they're moving to the latest Xeons.
I'm not kidding, I've actually seen a press release like this.
Now i'll read the article ...
EDIT: I would say i'm no more than 30% right. They were doing heavyweight data crunching in the cloud, and so paying more for it than if they were doing it on rented hardware. But that's a constant-factor thing; it's not like they were downloading gigabytes of CSVs from S3 on every request or some such. Their query looks suspect to me: couldn't it be written to do one big scan, rather than unioning a load of things? Or is this the right way to write queries on column stores? Still, there is no glaring obvious (to me) old-school fix for this.
> Refreshing caches automatically
How do people usually handle this? Is this something done on the application layer or database layer? Where is the cache stored?
Expiry can take a few different forms. Some caches have a defined space and use a replacement scheme like "fill the cache up, then remove the least recently accessed value". Some don't have defined sizes but instead remove entries based on timestamps (cache for n minutes). Some depend on invalidation messages from the application. It all depends on the applications needs.
The most important thing to remember is that caching means your system becomes inherently a distributed one. State can become split across multiple sources, the cache can return stale data, invalidation might not happen when you expect, ...
That's fine, but you have to program accordingly.
There seems to be a general opinion trend that the queries generated by the group builder algorithm are very inefficient, that it'd be easy to come up with a solution with much better response times, and that that would be achievable in any reasonable programming language in roughly the same time with similar results.
The language argument will always be controversial and I won't address it here; we have a point of view that is expressed in the Conclusion and on this blogpost: https://movio.co/en/blog/migrate-Scala-to-Go/
I can imagine that seeing a query with JOINs, subqueries, GROUP BYs and UNIONs can raise some eyebrows, but there is some lacking context in that story, and that's on me. Here's some of that context:
* The schema that the group builder algorithm operates on is not uniform in nature or composed of simple yes/no fields; it's an incredibly complex legacy schema that to a large degree wasn't even up to Movio: it's been up to the film industry as a whole, and it has evolved over the years, as is the case everywhere. Note that every different kind of filter translates to a very different kind of query, and we have more than 120 different filters, sometimes with dynamic parameters, and sometimes even bespoke for a particular customer!
* The group builder algorithm predates the team that built this service (myself included), as well as predating the first commercial release of Elasticsearch, MariaDB, mainstream Go success, etc. Nevertheless, it's still very fast and is being used today by ~88% of our customers (i.e. all the non-behemoths). It's been successful for many years, and continues to be, for the most part.
* But I don't like it because it's fast: I like it because it's simple and flexible. It allows our customers to build a really complex (and arbitrary) tree of filters to segment their loyalty member base, and it compiles all of that into one big SQL query, that in most cases is quite performant. That's pretty awesome. But yes; it doesn't scale to several million members.
* Migrating the very engine of the main product of a company is not a decision that is taken lightly. As is the case with every big company I can remember (e.g. Twitter, SoundCloud), behind a big success story there's always a legacy monolith, and our case is no exception. From that standpoint, achieving such breakthrough (i.e. cost reduction + significant response time improvement) within one hackathon day is really not all that common in my experience. Definitely something worth sharing, IMO.
Hopefully that clarifies some of the questions :) Cheers.
So if the alternative was Scala I can see why Go may have helped tighten things up a bit.
That's a different use case c.f. Facebook Ads.