Dashtable is very important data structure in Dragonfly. This document explain
how it fits inside the engine.
Each selectable database holds a primary dashtable that contains all its entries. Another instance of Dashtable holds an optional expiry information, for keys that have TTL expiry on them. Dashtable is equivalent to Redis dictionary but have some wonderful properties that make Dragonfly memory efficient in various situations.
## Redis dictionary
*“All problems in computer science can be solved by another level of indirection”*
This section is a brief refresher of how redis dictionary (RD) is implemented.
We shamelessly "borrowed" a diagram from [this blogpost](https://codeburst.io/a-closer-look-at-redis-dictionary-implementation-internals-3fd815aae535), so if you want a deep-dive, you can read the original article.
Each `RD` is in fact two hash-tables (see `ht` field in the diagram below). The second instance is used for incremental resizes of the dictionary.
Each hash-table `dictht` is implemented as a [classic hashtable with separate chaining](https://en.wikipedia.org/wiki/Hash_table#Separate_chaining). `dictEntry` is the link-list entry that wraps each key/value pair inside the table. Each dictEntry has three pointers and takes up 24 bytes of space. The bucket array of `dictht` is resized at powers of two, so usually its utilization is in [50, 100] range.
*Case 1*: it has `N` items at 100% load factor, in other words, buckets count equals to number of items. Each bucket holds a pointer to dictEntry, i.e. it's 8 bytes. In total we need: $8N + 24N = 32N$ bytes per record. <br>
*Case 2*: `N` items at 75% load factor, in other words, the number of buckets is 1.33 higher than number of items. In total we need: $N\*1.33\*8 + 24N \approx 34N$ bytes per record. <br>
*Case 3*: `N` items at 50% load factor, say right after table growth. Number of buckets is twice the number of items, hence we need $N\*2\*8 + 24N = 40N$ bytes per record.
Now lets take incremental growth into account. When `ht[0]` is full (i.e. RD needs to migrate data to a bigger table), it will instantiate a second temporary instance `ht[1]` that will hold additional 2*N buckets. Both instances will live in parallel until all data is migrated to `ht[1]` and then `ht[0]` bucket array will be deleted. All this complexity is hidden from a user by well engineered API of RD. Lets combine case 3 and case 1 to analyze memory spike at this point: `ht[0]` holds `N` items and it is fully utilized. `ht[1]` is allocated with `2N` buckets.
Overall, the memory needed during the spike is $32N + 16N=48N$ bytes.
[Dashtable](https://arxiv.org/abs/2003.07302) is an evolution of an algorithm from 1979 called [extendible hashing](https://en.wikipedia.org/wiki/Extendible_hashing).
Similarly to a classic hashtable, dashtable (DT) also holds an array of pointers at front. However, unlike with classic tables, it points to `segments` and not to linked lists of items. Each `segment` is, in fact, a mini-hashtable of constant size. The front array of pointers to segments is called `directory`. Similarly to a classic table, when an item is inserted into a DT, it first determines the destination segment based on item's hashvalue. The segment is implemented as a hashtable with open-addressed hashing scheme and as I said - constant in size. Once segment is determined, the item inserted into one of its buckets. If an item was successfully inserted, we finished, otherwise, the segment is "full" and needs splitting. The DT splits the contents of a full segment in two segments, and the additional segment is added to the directory. Then it tries to reinsert the item again. To summarize, the classic chaining hash-table is built upon a dynamic array of linked-lists while dashtable is more like a dynamic array of flat hash-tables of constant size.
In the diagram above you can see how dashtable looks like. Each segment is comprised of `K` buckets. For example, in our implementation a dashtable has 60 buckets per segment (it's a compile-time parameter that can be configured).
Below you can see the diagram of a segment. It comprised of regular buckets and stash buckets. Each bucket has `k` slots and each slot can host a key-value record.
In our implementation, each segment has 56 regular buckets, 4 stash buckets and each bucket contains 14 slots. Overall, each dashtable segment has capacity to host 840 records. When an item is inserted into a segment, DT first determines its home bucket based on item's hash value. The home bucket is one of 56 regular buckets that reside in the table. Each bucket has 14 available slots and the item can reside in any free slot. If the home bucket is full,
then DT tries to insert to the regular bucket on the right. And if that bucket is also full,
it tries to insert into one of 4 stash buckets. These are kept deliberately aside to gather
spillovers from the regular buckets. The segment is "full" when the insertion fails, i.e. the home bucket and the neighbour bucket and all 4 stash buckets are full. Please note that segment is not necessary at full capacity, it can be that other buckets are not yet full, but unfortunately, that item can go only into these 6 buckets,
so the segment contents must be split. In case of split event, DT creates a new segment,
adds it to the directory and the items from the old segment partly moved to the new one,
and partly rebalanced within the old one. Only two segments are touched during the split event.
Now we can explain why seemingly similar data-structure has an advantage over a classic hashtable
in terms of memory and cpu.
1. Memory: we need `~N/840` entries or `8N/840` bytes in dashtable directory to host N items on average.
Basically, the overhead of directory almost disappears in DT. Say for 1M items we will
need ~1200 segments or 9600 bytes for the main array. That's in contrast to RD where
we will need a solid `8N` bucket array overhead - no matter what.
For 1M items, it will obviously be 8MB. In addition, dash segments use open addressing collision
scheme with probing, that means that they do not need anything like `dictEntry`.
Dashtable uses lots of tricks to make its own metadata small. In our implementation,
the average `tax` per entry is short of 20 bits compared to 64 bits in RD (dictEntry.next).
In addition, DT incremental resize does not allocate a bigger table - instead
it adds a single segment per split event. Assuming that key/pair entry is two 8
byte pointers like in RD, then DT requires $16N + (8N/840) + 2.5N + O(1) \approx 19N$
bytes at 100% utilization. This number is very close to the optimum of 16 bytes.
In unlikely case when all segments just doubled in size, i.e.
DT is at 50% of utilization we may need $38N$ bytes per item.
In practice, each segment grows independently from others,
so the table has smooth memory usage of 22-32 bytes per item or **6-16 bytes overhead**.
1. Speed: RD requires an allocation for dictEntry per insertion and deallocation per deletion. In addition, RD uses chaining, which is cache unfriendly on modern hardware. There is a consensus in engineering and research communities that classic chaining schemes are slower tha open addressing alternatives.
into a table's metadata in order to implement other intelligent algorithms like forkless save. This is where some the Dragonfly's disrupting qualities [can be seen](#forkless-save).
To compare RD vs DT I often use an internal debugging command "debug populate" that quickly fills both datastores with data. It just saves time and gives more consistent results compared to memtier_benchmark.
It also shows the raw speed at which each dictionary gets filled without intermediary factors like networking, parsing etc.
Due to shared-nothing architecture, Dragonfly maintains a dashtable per thread with its own slice of data. Each thread fills 1/8th of 20M range it owns - and it much faster, almost 8 times faster.You can see that the total usage is even smaller, because now we maintain
This example shows how much memory Dragonfly uses during BGSAVE under load compared to Redis. Btw, BGSAVE and SAVE in Dragonfly is the same procedure because it's implemented using fully asynchronous algorithm that maintains point-in-time snapshot guarantees.
This test consists of 3 steps:
1. Execute `debug populate 5000000 key 1024` command on both servers to quickly fill them up
with ~5GB of data.
2. Run `memtier_benchmark --ratio 1:0 -n 600000 --threads=2 -c 20 --distinct-client-seed --key-prefix="key:" --hide-histogram --key-maximum=5000000 -d 1024` command in order to send constant update traffic. This traffic should not affect substantially the memory usage of both servers.
3. Finally, run `bgsave` on both servers while measuring their memory.
It's very hard, technically to measure exact memory usage of Redis during BGSAVE because it creates a child process that shares its parent memory in-part. We chose `cgroupsv2` as a tool to measure the memory. We put each server into a separate cgroup and we sampled `memory.current` attribute for each cgroup. Since a forked Redis process inherits the cgroup of the parent, we get an accurate estimation of their total memory usage. Although we did not need this for Dragonfly we applied the same approach for consistency.
As you can see on the graph, Redis uses 50% more memory even before BGSAVE starts. Around second 14, BGSAVE kicks off on both servers. Visually you can not see this event on Dragonfly graph, but it's seen very well on Redis graph. It took just few seconds for Dragonfly to finish its snapshot (again, not possible to see) and around second 20 Dragonfly is already behind BGSAVE. You can see a distinguishable cliff at second 39
that their their memory footprint could be reduced by as much as by 60% by employing better expiry methodology. The authors of the post above show pros and cons of expiration methods in the table below:
compartmentalized structure it can actually employ a very efficient passive expiry algorithm with low CPU overhead. Our passive procedure is complimented with proactive gradual scanning of the table in background.
We scan its buckets for the expired items. If we delete some of them, we may avoid growing the table altogether! The cost of scanning the segment before potential split is no more the
Please note that Redis could sustain 30% less qps. That means that the optimal working sets for Dragonfly and Redis are different - the former needed to host at least `20s*131k` items
