What is a better approach when it comes to performance and optimal resource utilization: accessing a database multiple times through AJAX to only get the exact information needed when it is needed, or performing one access to retrieve an object that holds all information that might be needed, with a high probability that not all is actually needed?

I know how to benchmark the actual queries, but I don't know how to test what is best when it comes to database performance when thousands of users are accessing the database simultaneously and how connection pooling comes into play.

  • whch platform ar u using.? if LAMP u cud use memcaching – ravi404 Dec 18 '12 at 21:48
  • Same as any other performance optimization, you measure it. – Telastyn Dec 18 '12 at 21:50
  • 2
    @Telastyn: I am making some fundamental design decissions and don't have a staging server. All my db-calls are to a a db that resides on the same machine where the php is executed. I was hoping to learn from other people's experience in this regard, before I come to the realization that the route I decided to take was great when everything was local, but sub-optimal when taken live. – DudeOnRock Dec 18 '12 at 21:57
  • 1
    @DudeOnRock - nod in general it depends on your usage patterns and how the data changes. If one query provides 80% of what people need and the data doesn't often change, then go with that. Easy to cache, easy to optimize. If one query returns like 5% of what users usually need, then maybe not. I would tend towards more queries than less. You can always cut them off at the server before it gets to the DB. Harder to undo 'everything makes one query'. – Telastyn Dec 18 '12 at 22:02
  • @ravz: sounds interesting! – DudeOnRock Dec 18 '12 at 22:02

There is no one correct answer to this; like any optimization it depends heavily on context / usage.

However, consider the following as a rule of thumb:

+: Data is stable / static
-: Data is dynamic / volatile

+: Data is frequently used
-: Data is infrequently used

++: fetch large chunks in the fewest number of fetches 
    and persist the data as long as possible within tolerances for staleness.

+-: do what is expedient to the logic & usage; if it is convenient to 
    fetch / calc as needed do so, if it is convenient to pre-fetch and 
    persist then do so. Seek to optimize only if absolutely necessary.

-+: fetch / calc as needed; but if optimization is required consider 
    pre-fetching or pre-calculating if possible, or negotiate a tolerance 
    for less than real time accuracy to reduce volatility.

--: fetch / calc as needed and don't worry about it further unless a 
    specific case is unacceptably expensive; if so see -+.
| improve this answer | |

Remember the first rule of optimization: measure, don't guess. Try both, instrument them with some sort of stopwatch code, and see what takes longer.

And also bear in mind the old joke that "there are only two difficult problems in computer science: cache invalidation and naming things well." If you pull everything from the DB at once and hold it in memory, you have a cache. And now you have a new problem: anytime anything changes anywhere in the system, it has to make the same change in two places: the database, and the cache. If you have more than one server talking to the DB, or multiple APIs to make the server modify data, this can become very tricky very quickly.

| improve this answer | |
  • And be sure what you measure. For example results can vary depending on the database connection bandwith and latency. – SpaceTrucker Dec 21 '12 at 13:44

There is NO silver bullet solution to this question. I guess you need to TRY the possible trade-offs and tune-up your server(s) to achieve the best out of it.

First point: before starting to do any improvements you need to SET your current performance benchmark, measuring it, and take it a baseline in comparison possible solutions to improve it.

The Second thing is, application usage needs to be tracked. The way how the application is utilized by end users. Cutting down on returned data raw numbers that is not needed to the end user(s) may save you a lot precious server resources. For example: there is no point in returning 5000 records while users are interested in first 50.

Third point: You have to understand the frequency of calls and possible implications. For example: if most calls are look-up value table queries, then you might probably would create an infrastructure to cache these calls. In another words, if your data is not changing frequently, consider the caching option. And of course, minimizing the number of calls always should help to boost performance.

| improve this answer | |

Getting everything at once will give you better performance, unless "everything" includes things like BLOBs or similarly large data objects. The performance overhead to serialize everything, move it over the wire, then deserialize it on the other end is pretty significant, with network latency being a big chunk of it. Memory is cheaper than network bandwidth, and will probably remain so for awhile yet. Your only real answer will come from a benchmark, but if you're just trying to gauge one over the other, that's the way I'd lean.

| improve this answer | |
  • According to the comments, this is using a local database, so there's no "over the wire" latency here. – Mason Wheeler Dec 18 '12 at 23:39
  • 1
    According to the comments, he was looking for strategies that wouldn't be "great when everything was local, but sub-optimal when taken live". – TMN Dec 19 '12 at 13:27

If you're making architectural decision, then REST is one option. With REST, you always request for a resource multiple times, i.e. you don't send a request to get 2 objects because each object has its own url. The performance concern with doing it this style will probably be solved when HTTP/2.0 come out. Otherwise, you just optimise to make it as fast as possible. A lot of companies are doing it this way.

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.