How does a B-tree make queries fast?
blog.allegro.tech
blog.allegro.tech
This all means we can implement indexes as b-trees where the keys are the values of a particular column and the value is the file offset of the row with that value. You could envision a simple db format where indexes and the main row file are stored in separate files. In such a database you could drop an index simply by deleting the indexes file (or add one by creating it). The main row file actually has all of the data and so indexes can be recreated if necessary (at expense of course).
The crux is understanding what data access patterns you will have and what indexes / data structures accelerate that access pattern. "Index = fast" is a painfully pernicious untrue meme. It's absolutely true for application tables with queries only touching a few rows. On the other hand, analytics queries touching a high proportion of rows with joins on equality conditions (ie. hash joinable) isn't going to go any faster with an index.
I've seen devs shotgun indexes at tables to fix performance (done it myself too) but the real test of index understanding is when that doesn't work.
That's when you bring a BRIN to the table. :-)
All databases are datastructures on disk and memory optimized for specific access and write patterns.
I believe that lack of internalization of that meme (regardless of how true it is) can be a cause of real trouble.
I was working in a team where Java devs simply didn't bother to put indexes on tables because they were small (like 100 rows or so). When I (JS dev) pestered them long enough to finally do it suddenly the whole app got super snappy and they were very thankful as it happened just as degrading performance was causing a lot of gloomy mood.
Why are you mentioning that it were Java devs ad that you are a JS (javascript?) dev? Does that give you any kind of... expertise in the matter?
With correct indexes the queries were able to be found only with index lookup without touching data at all. It was many times faster than sequential scan of even this few rows.
I'm mentioning our respective roles to show that people with nominally no expertise make such decisions often in practice and such "memes", even tough they are not strictly true in all cases, may bring a lot of value in practical setting.
"No inedex = slow" is a good heuristic for nearly all devs.
That doesn't make things faster, unless the rows contain so much data, that they force many page reads. Otherwise, reading the whole index or reading all rows takes the same time.
I think there is some information that you are not providing that might make a difference - or it is actually a misunderstanding on your side and if the queries really got faster, it's because of something else, such as a join.
The only thing I'm not providing are the exact sizes of the tables involved other than that they were considered small. Some having 100 rows but I can't rule out that some were 1000. I didn't say each of those tables fit a single page which I agree could be significant. I am completely sure that the difference was caused by adding indexes to small tables, nothing else. The improvement was achieved in a single day by a sigle person adding indexes where there were none. He observed improvement as he went from table to table. I was specifically thanked by him for pestering him about it. He was informal tech lead for this project. The application was fairly large at that time and performance issues were everywhere. What was also used everywhere were those small tables containing stuff like list of countries, languages, currencies, types of operations and such.
Depends on the data, and the index. A covering index would almost certainly be faster. But you said analytics, which implies large results from huge datasets, so it’s unlikely to fit into memory in the first place.
We have a winner. But when looking at SQL tables/Views/Stored Procedures, the data is also stored in order in memory, in effect have a master database and files on disk, with sorted databases and files in memory for faster access.
No it’s not… if all you do is write to it. In fact, it’s the fastest possible database for such case.
Indexes are pure redundancy - they contain the data already in the base table which must be maintained during writes.
But they can make reads so much faster, if the data access pattern can utilize them. The key is to identify access patterns which justify the price of the index.
> /dev/null will suffice.
And if you ever need to read anything even once database without indexes is slow
"database = fast" is way worse heuristic to believe in for the people that need heuristics to move on with what they are doing.
This may be true in most cases, but not all. It just depends on your access pattern. If all you do are full table scans (possibly feeding to hash-joins), you won't benefit from indexes at all!
The whole point is that indexes do not auto-magically improve performance with no downsides. If they did, we could just index all columns and call it a day!
In most cases it means that you shouldn't be doing what youa re trying to do. Or best case, that maybe database is not the right tool for your job.
The other direction, however, is not a given: most DBs will let you create a FK on fields not covered by an index, so deleting or modifying a parent can benefit if you create such index explicitly, because it can check for the existence of children much faster (and avoid potentially locking the entire table). Again, the access pattern governs what indexes are needed: if you never delete/modify parent, you may not need an index on FK (unless you also have some queries which can use it, of course).
then …
> the kernel seemed to struggle …
What was the struggle? If it’s performance doesn’t that contradict your earlier statement?
Genuinely interested in what the issue was, not trying to be a pedant
Performance is always great until you have to hit disk. Not uncommon to rely on mmap at which point your disk access is sub-optimal vs. a hand-tailored buffer manager with strategies to improve disk reads.
You don't need to use direct I/O to have fine control.
A custom page manager would have probably done the trick, but I don't have the technical chops to write one.
They already envisioned it in DBF and CDX.
(Here discussed in terms of skip lists, but they are similar enough that the distinction doesn't matter)
For B-trees it's only going to be a significant win if the chance of intersection is small, less than about `1/(nodes_per_block)`. For binary trees it's a much bigger win since that becomes 1/2 and binary trees are horrible on cache.
Hmm, can you efficiently intersect an arbitrary number of B-trees using the same idea as a heap-merge, but with a max heap instead of a min heap? You'd still have to iterate over all on a match, but as long as 2 input trees don't have an intersection, you don't have to look at the others at all ... or does that become equivalent to just doing them in series?
I guess there are worst case scenarios with evenly spaced intersections spread out exactly one per block, but in terms of block reads it fundamentally doesn't matter how you intersect two such trees, you'll still have to read all blocks, and that is orders of magnitude slower than comparing the values within.
I think the tree structure can be considered cached in real world scenarios; not really relevant to the performance you'll get.
For spinning disks, this is obvious. But why is sequential access faster on SSDs?
In an SSD, a write operation can only be done when the page is already erased. However, the unit of read/write operations are a page, while the unit of erase operation is a block. That means for a disk write, a naive implementation needs to read the whole block, erase the block, then write updated data back to the block, which is unacceptable. Furthermore, blocks should wear out uniformly, otherwise, the SSD would lose capacity.
To tackle these problems, SSD introduces Flash Translation Layer (FTL) which helps to build an illusion of random access device. To achieve this, FTL employs an approach very similar to LSM trees. Writes are always written to new, already erased pages, while in the background, garbage collects (GC) outdated data. FTL needs to keep a map from the user’s logical address to physical address on SSD, both in-memory and persistently.
So to answer your question, why are sequential writes are faster than random writes on SSDs? Because the address map table is smaller since new data is consecutive in larger chunks. Garbage Collection is simpler and only metadata needs to be updated. Erasing a block is required anyway.
Though with mmap the OS may do a speculative async readahead, depending on memadvise.
IOW what makes it faster to access sequentially on SSD
I wonder how true this is for the top commercial engines (Oracle, MS, IBM, etc.) whose internals are closed source and proprietary. Even a decade ago my experience performance testing Exadata implied some exotic magic at work, ie lookups are way faster than the expected O(log n). More recently while testing SQL Server's ability to join hundreds of tables together the performance was _way_ in excess of what I expected. I can't imagine these things have internals all that similar to say the B+Tree inside MySQL for example.
If you sit down and think carefully about your data you can often make even a simple bare-bones B-tree perform fantastically for a query, well in excess of what you'd get out of mysql or sqlite (which are already pretty fast).
I've been playing around with an idea that involves putting sql server at the heart of a game engine, and it is turning into one of the biggest rabbit holes I've ever explored. I thought latency/jitter would be more of a problem but it simply isn't.
In memory during a query it can use temporary indexes of other types, primarily hash tables and bitmaps.
Its performance on ad-hoc complex queries is about as good as it gets, few if any other RDBMS can beat its performance, but under the hood it’s still mostly just doing joins on b-trees!
I could see the on-disk format needing to be simple and stable, but once the datas buffered who knows what structures and algorithms these proprietary engines are using? You would need to have done some reverse engineering or had hands-on details from the inside which presumably comes w/legal consequences for leaking them.
Generally the secret sauce in these things is the query optimiser heuristics.
The actual data structures and algorithm are often relatively simple.
Having said that, I’ve read their whitepaper on how they implement hash tables, and… it’s way more complex than I had assumed.
They cater for scenarios like many duplicated keys, parallel construction, unbalanced load across CPU cores, etc…
But in any case, both of the varieties mentioned still have random components; they aren’t purely sequential. Monotonic for the time component, yes.
https://www.cybertec-postgresql.com/en/unexpected-downsides-...
This makes a big difference for the primary key (clustered index for both MySQL and postgresql).
Does anyone have an analysis for MySQL secondary indexes like the above referenced one for PostgreSQL secondary indexes?
In practice if you want good performance and scalability it's important to select keys well.
However uuid4 create a lot of work during updates, and tend to result in relatively low fill trees so querying is less efficient than it could be.
They also don’t really benefit from btrees as range or prefix queries are extremely rare.