Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Use Concurrency To Download Vendor List #4005

Open
wants to merge 5 commits into
base: master
Choose a base branch
from

Conversation

zhongshixi
Copy link
Contributor

@zhongshixi zhongshixi commented Oct 22, 2024

Problem
The current implementation of downloading vendor list in preload stage is executed sequentially for individual vendor file - it costs 5 to 10 seconds to download all ~300 vendor files on my local machine ( Apple M1 Pro, 32GB memory).

Solution

  1. the solution is to optimize the preload procedure by applying concurrency which is configurable by the end user. The end user can tune the number depending on their own spec of the machine.

  2. by default, the concurrency is disabled so it wont impact anything. It is end-user's responsibility to provide a number to tune the performance.

  3. in my pbs.yaml config , i set the following value which i observed almost 5x to 10x performance improvement ( ~500 to 600 ms ), any number more than 20 for max_concurrency_init_fetch_specific_version , the performance increase is diminishing

gdpr:
  default_value: "0" # disable gdpr, explicitly specifying a default value is a requirement in prebid server config
  vendorlist_fetcher:
    max_concurrency_init_fetch_latest_version: 2
    max_concurrency_init_fetch_specific_version: 20
  timeouts_ms:
    init_vendorlist_fetches: 30000
    active_vendorlist_fetch: 30000

you are welcome to play around with the solution on your own setting and i highly recommend you see the number yourself.

…ncurrency behaviour; add sample configuratio in the sample
@@ -284,12 +292,6 @@ var vendorList2Expected = testExpected{
vendorPurposes: map[int]bool{1: false, 2: true, 3: true},
}

var vendorListFallbackExpected = testExpected{
Copy link
Contributor Author

@zhongshixi zhongshixi Oct 22, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

this code block is not used at all so i deleted it

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Did you need to delete it to accomplish your goal in this PR ?

Copy link
Contributor Author

@zhongshixi zhongshixi Oct 22, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

but why we need to keep it if the code is not used at all? there is no comment or description associated with the deleted test case.

i used a practice called grooming as you go - so if you happen to touch the related code piece as part of the solution, then you just do a little bit grooming to avoid large bulk of grooming/refactoring in the future

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ok. I see no problem with that.

config/config.go Outdated
@@ -240,10 +240,21 @@ type Privacy struct {
LMT LMT
}

type VendorListFetcher struct {
// Max concurrent request for downloading the latest version of the vendor list from a specific major version.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think that the word "request" should be spelled in the plural form.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

good catch

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

fixed

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks !

config/config.go Outdated
// If it is 0 or negative, then means there is no limit for the max concurrency.
MaxConcurrencyInitFetchLatestVersion int `mapstructure:"max_concurrency_init_fetch_latest_version"`

// Max concurrent request for downloading every specific version of the vendor list from its first version to its latest version.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think that the word "request" should be spelled in the plural form.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

good catch

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

fixed

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks !

Comment on lines -244 to +256
Enabled bool `mapstructure:"enabled"`
HostVendorID int `mapstructure:"host_vendor_id"`
DefaultValue string `mapstructure:"default_value"`
Enabled bool `mapstructure:"enabled"`
HostVendorID int `mapstructure:"host_vendor_id"`
DefaultValue string `mapstructure:"default_value"`
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think that the vertical alignment was intentional.

Why change the indentation style of those lines ?

Copy link
Contributor Author

@zhongshixi zhongshixi Oct 22, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@sebhtml do you mind elaborating more about vertical alignment was intentional ? Changing the indentation is automatically done by the go formatter ( i am using MS Visual Studio ) , which i believe it is easier to read.

but i am curious about other technical aspects related to it

Copy link
Contributor

@sebhtml sebhtml Oct 22, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The types are vertically aligned:

                                |
                                v
	Timeouts                GDPRTimeouts `mapstructure:"timeouts_ms"`
	NonStandardPublishers   []string     `mapstructure:"non_standard_publishers,flow"`
	NonStandardPublisherMap map[string]struct{}
	TCF2                    TCF2 `mapstructure:"tcf2"`
	AMPException            bool `mapstructure:"amp_exception"`

see

Timeouts GDPRTimeouts `mapstructure:"timeouts_ms"`

I don't know if my comment is relevant though.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It may be go fmt's decision; they actually look wrong before.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@bsardo let me know your opinion

config/config.go Outdated
Comment on lines 1157 to 1158
v.SetDefault("gdpr.vendorlist_fetcher.max_concurrency_init_fetch_latest_version", 1) // by default one download at a time
v.SetDefault("gdpr.vendorlist_fetcher.max_concurrency_init_fetch_latest_version", 1) // by default one download at a time
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why is this line repeated twice ?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

good catch, will change it

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

fixed

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks !

Comment on lines 92 to 110
tsStart := time.Now() // For logging how long this takes
for _, v := range versions {
specVersion := v.specVersion
firstVersion := v.firstListVersion
wgLatestVersion.Go(func() error {
latestVersion := saveOne(ctx, client, urlMaker(specVersion, 0), saver)
for i := firstVersion; i < latestVersion; i++ {
currentVersion := i
wgSpecificVersion.Go(func() error {
saveOne(ctx, client, urlMaker(specVersion, currentVersion), saver)
return nil
})
}
return nil
})
}

wgLatestVersion.Wait()
wgSpecificVersion.Wait()
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi !

I don't know if it matters.

The WaitGroup wgSpecificVersion will likely wait for more Goroutines than wgLatestVersion.

Globally, you start Goroutines with a limit of conf.MaxConcurrencyInitFetchLatestVersion + conf.MaxConcurrencyInitFetchSpecificVersion concurrent Go routines.

So why have 2 wait groups instead of just 1 wait group ?

Copy link
Contributor Author

@zhongshixi zhongshixi Oct 22, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@sebhtml , good thought, however the reason i did it is to have fine granular control of two different behaviours:

  1. fetching the latest version
  2. fetching each specific version.

plus provide sequential execution as default behaviour so the decision of tuning the controlling parameters is determined by the host.


by using only one wait group, there are two thoughts behind it

  1. conf.MaxConcurrencyInitFetchLatestVersion and conf.MaxConcurrencyInitFetchSpecificVersion loses its meaning since one wait group wont be able to control two types of behaviour precisely.

  2. lets say you use one wait group and the corresponding config is conf.MaxConcurrencyInitFetch , and then you set the value to 1 ( which is supposed to mean it is a sequential execution ) with the following implementation

var wg errorgroup.Group
for _, v := range versions {
		specVersion := v.specVersion
		firstVersion := v.firstListVersion
		wg.Go(func() error {
			latestVersion := saveOne(ctx, client, urlMaker(specVersion, 0), saver)
			for i := firstVersion; i < latestVersion; i++ {
				currentVersion := i
				wg.Go(func() error {
					saveOne(ctx, client, urlMaker(specVersion, currentVersion), saver)
					return nil
				})
			}
			return nil
		})
	}

then you have deadlock because one go routine is waiting for another go routine while only 1 go routine is available

let me know if you have different idea

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Oh I see. You have a good point. I therefore resolve this thread.

Comment on lines 203 to 205
saverMutex.Lock()
*s = append(*s, vi)
saverMutex.Unlock()
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If *s = append(*s, vi) panics, I think that the saverMutex will never be unlocked.

I think that deferring can avoid that:

	saverMutex.Lock()
	defer saverMutex.Unlock()
	*s = append(*s, vi)

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

updated

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks

Comment on lines +9 to +14
vendorlist_fetcher:
max_concurrency_init_fetch_latest_version: 2
max_concurrency_init_fetch_specific_version: 20
timeouts_ms:
init_vendorlist_fetches: 30000
active_vendorlist_fetch: 30000
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Are those new configuration keys optional ?

Copy link
Contributor Author

@zhongshixi zhongshixi Oct 22, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

if init_vendorlist_fetches is 0 - then you literally set 0 sec as the timeout of preloading the vendor list, what you end up seeing the preload will have context deadline exceeded/timeout

active_vendorlist_fetch is more for fetching the vendor list on the run - same idea as above

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi @zhongshixi

I am not asking what are those new configuration keys for.

I am asking whether or not the Prebid Server software works if you don't provide them.

Copy link
Contributor Author

@zhongshixi zhongshixi Oct 22, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@sebhtml
for vendor list max concurrent request

	v.SetDefault("gdpr.vendorlist_fetcher.max_concurrency_init_fetch_latest_version", 1)   // by default one fetch at a time
	v.SetDefault("gdpr.vendorlist_fetcher.max_concurrency_init_fetch_specific_version", 1) // by default one fetch at a time

we have default settings, which is literally the same as before ( sequential download )

for init_vendorlist_fetches and active_vendorlist_fetch, here is the default, if you do not set them to a reason value, you wont be able to preload the vendor list due to context deadline issue. The server can still spin up and run perfectly, but you do not have a prefetched vendor list

v.SetDefault("gdpr.timeouts_ms.init_vendorlist_fetches", 0)
	v.SetDefault("gdpr.timeouts_ms.active_vendorlist_fetch", 0)

let me know if i answered your question

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

So it's not backward-compatible.
I think that it should be backward-compatible.

Copy link
Contributor Author

@zhongshixi zhongshixi Oct 22, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@sebhtml i am not following

  1. even without my change the issue context deadline issue still happen, because you can not really set 0 timeout , the default setting is using 0 as timeout, i talked to many people who host prebid server, they all configure the following value to more than 0 so we do not immediately timeout when the fetchiing request is fired
v.SetDefault("gdpr.timeouts_ms.init_vendorlist_fetches", 0)
v.SetDefault("gdpr.timeouts_ms.active_vendorlist_fetch", 0)
  1. the default logic before my PR change uses sequential execution, the default setting added in this PR also leads to sequential execution
v.SetDefault("gdpr.vendorlist_fetcher.max_concurrency_init_fetch_latest_version", 1)   // by default one fetch at a time
	v.SetDefault("gdpr.vendorlist_fetcher.max_concurrency_init_fetch_specific_version", 1) // by default one fetch at a time

in summary, i do not think this PR incurs any compatibility issue. So i would like to ask if could you point out exactly which part of the code logic is not backwards compatible, maybe i am actually missing something

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't know how to rephrase it. So I let it go.

@bsardo bsardo self-assigned this Oct 22, 2024
for i := firstVersion; i < latestVersion; i++ {
currentVersion := i
wgSpecificVersion.Go(func() error {
saveOne(ctx, client, urlMaker(specVersion, currentVersion), saver)
Copy link
Contributor Author

@zhongshixi zhongshixi Oct 28, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

to help reviewer understand the reason behind reason that i do not use channel or wait group+mutex is because the implementation of saveOne is already mutex safe due to its using sync.Map underneath, there is no need to add unnecessary new lock mechanism

https://pkg.go.dev/sync#Map

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

4 participants