Lets take an example of Twitter. There is a huge cache which gets updated frequently. For example: if person Foo tweets and it has followers all across the globe. Ideally all the caches across all PoP needs to get updated. i.e. they should remain in sync
How does replication across datacenter (PoP) work for realtime caches ?
What tools/technologies are preferred ?
What are potential issues here in this system design ?
I am not sure there is a right/wrong answer to this, but here's my two pennies' worth of it.
I would tackle the problem from a slightly different angle: when a user posts something, that something goes in a distributed storage (not necessarily a cache) that is already redundant across multiple geographies. I would also presume that, in the interest of performance, these nodes are eventually consistent.
Now the caching. I would not design a system that takes care of synchronising all the caches each time someone does something. I would rather implement caching at the service level. Imagine a small service residing in a geographically distributed cluster. Each time a user tries to fetch data, the service checks its local cache - if it is a miss, it reads the tweets from the storage and puts a portion of them in a cache (subject to eviction policies). All subsequent accesses, if any, would be cached at a local level.
In terms of design precautions:
Carefully consider the DC / AZ topology in order to ensure sufficient bandwidth and low latency
Cache at the local level in order to avoid useless network trips
Cache updates don't happen from the centre to the periphery; cache is created when a cache miss happens
I am stating the obvious here, implement the right eviction policies in order to keep only the right objects in cache
The only message that should go from the centre to the periphery is a cache flush broadcast (tell all the nodes to get rid of their cache)
I am certainly missing many other things here, but hopefully this is good food for thought.
Related
I was pondering this question today. An LRU cache in the context of a database in a web app helps ensure Availability with fast data lookups that do not rely on continually accessing the database.
However, how does an LRU cache in practice stay fresh? As I understand it, one cannot garuntee Consistency along with Availibility. How is a frequently used item, which therefore does not expire from the LRU cache, handle modification? Is this an example where in a system that needs C over A, an LRU cache is not a good choice?
First of all, a cache too small to hold all the data (where an eviction might happen and the LRU part is relevant) is not a good example for the CAP theorem, because even without looking at consistency, it can't even deliver partition tolerance and availability at the same time. If the data the client asks for is not in the cache, and a network partition prevents the cache from getting the data from the primary database in time, then it simply can't give the client any answer on time.
If we only talk about data actually in the cache, we might somewhat awkwardly apply the CAP-theorem only to that data. Then it depends on how exactly that cache is used.
A lot of caching happens on the same machine that also has the authoritative data. For example, your database management system (say PostgreSql or whatever) probably caches lots of data in RAM and answers queries from there rather than from the persistent data on disk. Even then cache invalidation is a hairy problem. Basically even without a network you either are OK with sometimes using outdated information (basically sacrificing consistency) or the caching system needs to know about data changes and act on that and that can get very complicated. Still, the CAP theorem simply doesn't apply, because there is no distribution. Or if you want to look at it very pedantically (not the usual way of putting it) the bus the various parts of one computer use to communicate is not partition tolerant (the third leg of the CAP theorem). Put more simply: If the parts of your computer can't talk to one another the computer will crash.
So CAP-wise the interesting case is having the primary database and the cache on separate machines connected by an unreliable network. In that case there are two basic possibilities: (1) The caching server might answer requests without asking the primary database if its data is still valid, or (2) it might check with the primary database on every request. (1) means consistency is sacrificed. If its (2), there is a problem the cache's design must deal with: What should the cache tell the client if it doesn't get the primary database's answer on time (because of a partition, that is some networking problem)? In that case there are basically only two possibilities: It might still respond with the cached data, taking the risk that it might have become invalid. This is sacrificing consistency. Or it may tell the client it can't answer right now. That is sacrificing availability.
So in summary
If everything happens on one machine the CAP theorem doesn't apply
If the data and the cache are connected by an unreliable network, that is not a good example of the CAP theorem, because you don't even get A&P even without C.
Still, the CAP theorem means you'll have to sacrifice C or even more of A&P than the part a cache won't deliver in the first place.
What exactly you end up sacrificing depends on how exactly the cache is used.
Assume some distributed CRUD Service that uses a distributed cache that is not read-through (just some Key-Value store agnostic of DB). So there are n server nodes connected to m cache nodes (round-robin as routing). The cache is supposed to cache data stored in a DB layer.
So the default retrieval sequence seems to be:
check if data is in cache, if so return data
else fetch from DB
send data to cache (cache does eviction)
return data
The question is whether the individual service nodes can be smarter about what data to send to the cache, to reduce cache capacity costs (achieve similar hit ratio with less required cache storage space).
Given recent benchmarks on optimal eviction/admission strategies (in particular LFU), some new caches might not even store data if it is deemed too infrequently used, maybe application nodes can do some best-effort guess.
So my idea is that the individual service nodes could evaluate whether data that was fetched from a DB should be send to the distributed cache or not based on an algorithm like LFU, thus reducing the network traffic between service and cache. I am thinking about local checks (suffering a lack of effectivity on cold startups), but checks against a shared list of cached keys may also be considered.
So the sequence would be
check if data is in cache, if so return data
else fetch from DB
check if data key is frequently used
if yes, send data to cache (cache does eviction). Else not.
return data
Is this possible, reasonable, has it already been done?
It is common in databases, search, and analytical products to guard their LRU caches with filters to avoid pollution caused by scans. For example see Postgres' Buffer Ring Replacement Strategy and ElasticSearch's filter cache. These are admission policies detached from the cache itself, which could be replaced if their caching algorithm was more intelligent. It sounds like your idea is similar, except a distributed version.
Most remote / distributed caches use classic eviction policies (LRU, LFU). That is okay because they are often excessively large, e.g. Twitter requires a 99.9% hit rate for their SLA targets. This means they likely won't drop recent items because the penalty is too high and oversize so that the victim is ancient.
However, that breaks down when batch jobs run and pollute the remote caching tier. In those cases, its not uncommon to see the cache population disabled to avoid impacting user requests. This is then a distributed variant of Postgres' problem described above.
The largest drawback with your idea is checking the item's popularity. This might be local only, which has a frequent cold start problem, or remote call which adds a network hop. That remote call would be cheaper than the traffic of shipping the item, but you are unlikely to be bandwidth limited. Likely you're goal would be to reduce capacity costs by a higher hit rate, but if your SLA requires a nearly perfect hit rate then you'll over provision anyway. It all depends on whether the gains by reducing cache-aside population operations are worth the implementation effort. I suspect that for most it hasn't been.
I am a software developer but wannabe architect new to the server scalability world.
In the context of multiple services working with the same data set, aiming to scale for redundancies and load balancing.
The question is: In a idealistic system, should services try to optimize their internal processing to reduce the amount of queries done to the remote server cache for better performance and less bandwidth at the cost of some local memory and code base or is it better to just go all-in and query the remote cache as the single transaction point every time any transaction need processing done on the data?
When I read about Redis and even general database usage online, the later seems to be the common option. Every nodes of the scaled application have no memory and read and write directly to the remote cache on every transactions.
But as a developer, I ask if this isn't a tremendous waste of resources? Whether you are designing at electronic chips level, at inter-thread, inter-process or inter-machine, I do believe it's the responsibility of each sub-system to do whatever it can to optimize its processing without depending on the external world if it can and hence reduce overall operation time.
I mean, if the same data is read over hundreds or time from the same service without changes (write), isn't it just more logical to keep a local cache and wait for notifications of changes (pub/sub) and only read only these changes to update the cache instead reading the bigger portion of data every time a transaction require it? On the other hand, I understand that this method implies that the same data will be duplicated at multiple place (more ram usage) and require some sort of expiration system not to keep the cache from filling up.
I know Redis is built to be fast. But however fast it is, in my opinion there's still a massive difference between reading directly from local memory versus querying an external service, transfer data over network, allocating memory, deserialize into proper objects and garbage collect it when you are finished with it. Anyone have benchmark numbers between in-process dictionaries query versus a Redis query on the localhost? Is it a negligible time in the bigger scheme of things or is it an important factor?
Now, I believe the real answer to my question until now is "it depends on your usage scenario", so let's elaborate:
Some of our services trigger actions on conditions of data change, others periodically crunch data, others periodically read new data from external network source and finally others are responsible to present data to users and let them trigger some actions and bring in new data. So it's a bit more complex than a single web pages deserving service. We already have a cache system codebase in most services, and we have a message broker system to notify data changes and trigger actions. Currently only one service of each type exist (not scaled). They transfer small volatile data over messages and bigger more persistent (changing less often) data over SQL. We are in process of moving pretty much all data to Redis to ease scalability and performances. Now some colleagues are having a heated discussion about whether we should abandon the cache system altogether and use Redis as the common global cache, or keep our notification/refresh system. We were wondering what the external world think about it. Thanks
(damn that's a lot of text)
I would favor utilizing in-process memory as much as possible. Any remote query introduces latency. You can use a hybrid approach and utilize in-process cache for speed (and it is MUCH faster) but put a significantly shorter TTL on it, and then once expired, reach further back to Redis.
I know that a big part of the performance from Couchbase comes from serving in-memory documents and for many of my data types that seems like an entirely reasonable aspiration but considering how user-data scales and is used I'm wondering if it's reasonable to plan for only a small percentage of the user documents to be in memory all of the time. I'm thinking maybe only 10-15% at any given time. Is this a reasonable assumption considering:
At any given time period there will be a only a fractional number of users will be using the system.
In this case, users only access there own data (or predominantly so)
Recently entered data is exponentially more likely to be viewed than historical user documents
UPDATE:
Some additional context:
Let's assume there's a user base of a 1 million customers, that 20% rarely if ever access the site, 40% access it once a week, and 40% access it every day.
At any given moment, only 5-10% of the user population would be logged in
When a user logs in they are like to re-query for certain documents in a single session (although the client does do some object caching to minimise this)
For any user, the most recent records are very active, the very old records very inactive
In summary, I would say of a majority of user-triggered transactional documents are queried quite infrequently but there are a core set -- records produced in the last 24-48 hours and relevant to the currently "logged in" group -- that would have significant benefits to being in-memory.
Two sub-questions are:
Is there a way to indicate a timestamp on a per-document basis to indicate it's need to be kept in memory?
How does couchbase overcome the growing list of document id's in-memory. It is my understanding that all ID's must always be in memory? isn't this too memory intensive for some apps?
First,one of the major benefits to CB is the fact that it is spread across multiple nodes. This also means your queries are spread across multiple nodes and you have a performance gain as a result (I know several other similar nosql spread across nodes - so maybe not relevant for your comparison?).
Next, I believe this question is a little bit too broad as I believe the answer will really depend on your usage. Does a given user only query his data one time, at random? If so, then according to you there will only be an in-memory benefit 10-15% of the time. If instead, once a user is on the site, they might query their data multiple times, there is a definite performance benefit.
Regardless, Couchbase has pretty fast disk-access performance, particularly on SSDs, so it probably doesn't make much difference either way, but again without specifics there is no way to be sure. If it's a relatively small document size, and if it involves a user waiting for one of them to load, then the user certainly will not notice a difference whether the document is loaded from RAM or disk.
Here is an interesting article on benchmarks for CB against similar nosql platforms.
Edit:
After reading your additional context, I think your scenario lines up pretty much exactly how Couchbase was designed to operate. From an eviction standpoint, CB keeps the newest and most-frequently accessed items in RAM. As RAM fills up with new and/or old items, oldest and least-frequently accessed are "evicted" to disk. This link from the Couchbase Manual explains more about how this works.
I think you are on the right track with Couchbase - in any regard, it's flexibility with scaling will easily allow you to tune the database to your application. I really don't think you can go wrong here.
Regarding your two questions:
Not in Couchbase 2.2
You should use relatively small document IDs. While it is true they are stored in RAM, if your document ids are small, your deployment is not "right-sized" if you are using a significant percentage of the available cluster RAM to store keys. This link talks about keys and gives details relevant to key size (e.g. 250-byte limit on size, metadata, etc.).
Basically what you are making a decision point on is sizing the Couchbase cluster for bucket RAM, and allowing a reduced residency ratio (% of document values in RAM), and using Cache Misses to pull from disk.
However, there are caveats in this scenario as well. You will basically also have relatively constant "cache eviction" where "not recently used" values are being removed from RAM cache as you pull cache missed documents from disk into RAM. This is because you will always be floating at the high water mark for the Bucket RAM quota. If you also simultaneously have a high write velocity (new/updated data) they will also need to be persisted. These two processes can compete for Disk I/O if the write velocity exceeds your capacity to evict/retrieve, and your SDK client will receive a Temporary OOM error if you actually cannot evict fast enough to open up RAM for new writes. As you scale horizontally, this becomes less likely as you have more Disk I/O capacity spread across more machines all simultaneously doing this process.
If when you say "queried" you mean querying indexes (i.e. Views), this is a separate data structure on disk that you would be querying and of course getting results back is not subject to eviction/NRU, but if you follow the View Query with a multi-get the above still applies. (Don't emit entire documents into your Index!)
For fun i am writing a fastcgi app. Right now all i do is generate a GUID and display it at the top of the page then make a db query based on the url which pulls data from one of my existing sites.
I would like to attempt to cache everything on the page except for the GUID. What is a good way of doing that? I heard of but never used redis. But it appears its a server which means its in a seperate process. Perhaps an in process solution would be faster? (unless its not?)
What is a good solution for page caching? (i'm using C++)
Your implementation sounds like you need a simple key-value caching mechanism, and you could possibly use a container like std::unordered_map from C++11, or its boost cousin, boost::unordered_map. unordered_map provides a hash table implementation. If you needed even higher performance at some point, you could also look at Boost.Intrusive which provides high performance, standard library-compatible containers.
If you roll your cache with the suggestions mentioned, a second concern will be expiring cache entries, because of the possibility your cached data will grow stale. I don't know what your data is like, but you can choose to implement a caching strategy like any of these:
after a certain time/number of uses, expire a cached entry
after a certain time/number of uses, expire the entire cache (extreme)
least-recently used - there's a stack overflow question concerning this: LRU cache design
Multithreaded/concurrent access may also be a concern, though as suggested in the link above, a possibility would be to lock the cache on access rather than worry about granular locking.
Now if you're talking about scaling, and moving up to multiple processes, and distributing server processes across multiple physical machines, the simple in-process caching might not be the way to go anymore (everyone could have different copies of data at any given time, inconsistency of performance if some server has cached data but others don't).
That's where Redis/Memcached/Membase/etc. shine - they are built for scaling and for offloading work from a database. They could be beaten out by a database and in-memory cache in performance (there is latency, after all, and a host of other factors), but when it comes to scaling, they are very useful and save load from a database, and can quickly serve requests. They also come with features cache expiration (implementations differ between them).
Best of all? They're easy to use and drop in. You don't have to choose redis/memcache from the outset, as caching itself is just an optimization and you can quickly replace the caching code with using, say, an in-memory cache of your own to using redis or something else.
There are still some differences between the caching servers though - membase and memcache distribute their data, while redis has master-slave replication.
For the record: I work in a company where we use memcached servers - we have several of them in the data center with the rest of our servers each having something like 16 GB of RAM allocated completely to cache.
edit:
And for speed comparisons, I'll adapt something from a Herb Sutter presentation I watched long ago:
process in-memory -> really fast
getting data from a local process in-memory data -> still really fast
data from local disk -> depends on your I/O device, SSD can be fast, but mechanical drives are glacial
getting data from remote process (in-memory data) -> fast-ish, and your cache servers better be close
getting data from remote process (disk) -> iceberg