Hello,
(cross-posting being something I want to get developers and users express their opinions).
We were recently in a (lucky) situation to have two new modules submissions targeting more or less the same purpose: allowing to use Kamailio with a Redis backend via database API.
One was submitted by Andreas Granig (Sipwise) and already merged with the name db_redis, because it was designed from the beginning as a generic DB connector, so the module can be used with auth_db, acc, usrloc, ...
* https://www.kamailio.org/docs/modules/devel/modules/db_redis.html
The second one was submitted by Surendra Tiwari (Plivo), initially having a naming conflict with db_redis, but renamed to db_redisusrloc, see the pull request at:
* https://github.com/kamailio/kamailio/pull/1446
Now, this email is about deciding the way to go forward with the second module.
It was designed to be used only for usrloc in the first phase, with many specific usrloc attributes hard coded inside db_redisusrloc. Surendra said (in a private chat) that the long term plan is to make it work for other modules. Anyhow, at this moment is very tied to usrloc, therefore the name of the module.
Given that the backend (Redis) is not an SQL engine, mapping over Kamailio's DB API needs some schema definition (see the readme of db_redis) in order to make it generic and work for all our modules that use a DB backend.
On the other hand, to squeeze the best of the backend, specially in no-SQL cases, having a dedicated DB connector module optimized for a specific module might help to get more performances/high-availability/scalability from the backend itself.
In this case, for example the expires value for a contact record can be set inside redis, so kamailio module doesn't have to run a timer routine to clean up (it doesn't mean db_redisusrloc does it right now, I didn't have time for a proper review, just giving an example). Surendra said they use it in production for couple of months now and it is several times faster than using usrloc with db_postgres (iirc, not db_mysql) for db_mode=3 (database only mode).
But of course, the reverse of the medal with a dedicated db connector for a module: it adds overhead to code maintenance (besides generic updates due to external library changes, I expect changing something relevant in usrloc, like adding new columns, would require updates in this module as well).
So, there are few things I want feedback on:
1) how do you fill about splitting from a generic-only DB connectors to have also some dedicated ones? This is more from confusion point of view, as a general rule so far, we do not deny contributions if there are other options for same kind of feature (e.g., many lcr or nat traversal options). As long as the contributor is willing to maintain the code, we were fine.
2) I guess usrloc, presence and dialog modules would be the main suspects that would benefit for such dedicated connectors, in other cases might not worth adding dedicated connectors. Missing any other module one would like to squeeze more performances with a dedicated connector?
3) should we set a different naming policy for such modules, for example: use *dbs_* prefix instead *db_*, to suggest better it is a DataBase Specific connector?
4) Andreas said he plans to do some performance testing of usrloc module with the two modules and see the differences. Anyone else that wants to do it? It can be a good metric to see if it worth going one way or another?
5) Helping to review the pull request, specially if you use Redis, is appreciated. Personally I am very short in available time these days, next week I plan do to new Kamailio stable releases, so the schedule is not getting lighter in my side.
Cheers, Daniel
Hi,
I have not dived into the technical specifics deeply or examined the innards of db_redis, so this is really shooting from the hip:
I can appreciate the very legitimate dilemma here. However, in general I would advocate for keeping all DB connectors generic.
Although I understand that there are some punishing generalisations and contortions required to shoehorn particular modules' schematic requirements into the more general DB API, there is a lot of value in the highly consistent database agnosticism afforded by doing this. This is similar to the trade-offs involved in using ORMs a lot of the time. As things currently stand, it is possible to use any db_* backend with Kamailio in a highly interchangeable way, and not many DB-backed systems can truly deliver on that promise. Kamailio shines as a rare exception.
The other problem is that the leaky / inconsistent abstractions involved in more Byzantine DB module choices will be confusing to users, most of whom are not software engineers presumably and will not easily understand the difference between a module-specific DB connector and a generic one. It will also lead to some needless duplication of effort and possibly lopsided support for one or the other pathway to the database in cases where a module has both generic and module-specific connector support.
Finally, although I understand that third parties contributing modules are charged with maintaining them, it seems to be a practical effect that you (Daniel) and others are dragged into maintaining orphaned or semi-orphaned modules, or the "plumbing" dependencies that they generate within common code. The project already has a bewildering array of modules, all moving at different speeds, having different levels of sophistication and so forth. Now, I know that I have always taken a more conservative position on matters of curation of third-party contributions relative to the implicit maintenance burden, and we may differ on that, but I think pouring more fuel on that fire by splicing in multiple levels of DB abstraction for some modules will create unnecessary maintenance headaches without the corresponding pay-off.
Whether the database is traditional relational or schemaless, I don't think it's unreasonable to ask that any DB-backed module adhere to the common DB API and use those mechanisms to interact with the DB. It just leads to more consistent, maintainable code and more predictable outcomes, while providing a level of flexibility people have come to expect with the RDBM-backed modules.
-- Alex
Hello,
for fairness and a more complete picture, I should have mentioned that I am somehow the first that broke into module specific backend with topos and topos_redis. In that case, I was also guided by the goals of squeezing what ever can be exploited in terms of performances from Redis backend, given that topology hiding with stripping/adding back headers for all SIP traffic needs be be (ideally) very fast.
The topos module had DB API support and can be used with mysql, postgres, etc ... topos_redis was added as an alternative to use directly a Redis server. However, topis_redis is not on top of DB API, therefore a name not related to it at all.
Cheers, Daniel
On 21.02.18 10:48, Alex Balashov wrote:
Hi,
I have not dived into the technical specifics deeply or examined the innards of db_redis, so this is really shooting from the hip:
I can appreciate the very legitimate dilemma here. However, in general I would advocate for keeping all DB connectors generic.
Although I understand that there are some punishing generalisations and contortions required to shoehorn particular modules' schematic requirements into the more general DB API, there is a lot of value in the highly consistent database agnosticism afforded by doing this. This is similar to the trade-offs involved in using ORMs a lot of the time. As things currently stand, it is possible to use any db_* backend with Kamailio in a highly interchangeable way, and not many DB-backed systems can truly deliver on that promise. Kamailio shines as a rare exception.
The other problem is that the leaky / inconsistent abstractions involved in more Byzantine DB module choices will be confusing to users, most of whom are not software engineers presumably and will not easily understand the difference between a module-specific DB connector and a generic one. It will also lead to some needless duplication of effort and possibly lopsided support for one or the other pathway to the database in cases where a module has both generic and module-specific connector support.
Finally, although I understand that third parties contributing modules are charged with maintaining them, it seems to be a practical effect that you (Daniel) and others are dragged into maintaining orphaned or semi-orphaned modules, or the "plumbing" dependencies that they generate within common code. The project already has a bewildering array of modules, all moving at different speeds, having different levels of sophistication and so forth. Now, I know that I have always taken a more conservative position on matters of curation of third-party contributions relative to the implicit maintenance burden, and we may differ on that, but I think pouring more fuel on that fire by splicing in multiple levels of DB abstraction for some modules will create unnecessary maintenance headaches without the corresponding pay-off.
Whether the database is traditional relational or schemaless, I don't think it's unreasonable to ask that any DB-backed module adhere to the common DB API and use those mechanisms to interact with the DB. It just leads to more consistent, maintainable code and more predictable outcomes, while providing a level of flexibility people have come to expect with the RDBM-backed modules.
-- Alex
Am Mittwoch, 21. Februar 2018, 10:33:23 CET schrieb Daniel-Constantin Mierla:
[..] We were recently in a (lucky) situation to have two new modules submissions targeting more or less the same purpose: allowing to use Kamailio with a Redis backend via database API. [..] So, there are few things I want feedback on:
- how do you fill about splitting from a generic-only DB connectors to
have also some dedicated ones? This is more from confusion point of view, as a general rule so far, we do not deny contributions if there are other options for same kind of feature (e.g., many lcr or nat traversal options). As long as the contributor is willing to maintain the code, we were fine.
- I guess usrloc, presence and dialog modules would be the main
suspects that would benefit for such dedicated connectors, in other cases might not worth adding dedicated connectors. Missing any other module one would like to squeeze more performances with a dedicated connector?
Hello,
I also think that the possibility to use a generic DB connection module to interact with several databases from different modules is really beneficial in Kamailio and we should keep this possibility.
Even if there is a small performance penalty in using a more generic module I still think the benefits in maintenance over the long term are in favor of only one generic redis module (as in this example). In the end processors are still getting faster/getting more cores every year. Our maintenance and reviewing capacity especially for core and plumping parts stays more or less the same.
If there are bigger performance differences (like an order of magnitude) between the two modules then the reason is probably some implementation issues in the code. In this case they should be simply fixed in the generic module.
Best regards,
Henning
- should we set a different naming policy for such modules, for
example: use *dbs_* prefix instead *db_*, to suggest better it is a DataBase Specific connector?
- Andreas said he plans to do some performance testing of usrloc module
with the two modules and see the differences. Anyone else that wants to do it? It can be a good metric to see if it worth going one way or another?
- Helping to review the pull request, specially if you use Redis, is
appreciated. Personally I am very short in available time these days, next week I plan do to new Kamailio stable releases, so the schedule is not getting lighter in my side.
Cheers, Daniel
Opinion of a Kamailio novice, take with a grain of salt:
Would it be appropriate to think about this general problem more akin to how some modules support either a file or a db backend? Imagine this world:
- To be called a DB in Kamailio parlance, a backend needs to conform to the DB API, be generically usable, etc – all the good stuff we have now. - Some module (foo) has 3 modes of operation: - file – the module uses a file on disk - db – the module uses the DB api - redis – a very foo-specialized application of Redis and its features
In my mind, having a specialized Redis backend for a module is no different than having a file backend in addition to a DB backend (see the dispatcher module for example).
Just because something (redis) is database-like, doesn't necessarily mean that it can/should only be accessed as a DB (above definition), especially if the tool supports features that shouldn't be in the DB API, and those features are very valuable to a specific module.
This would imply that there is a db_redis module which allows Redis to participate in anything demanding a DB (as defined above), as well as a foo_redis (name not important) module, which leverages Redis in a very foo-specific way.
It would of course be the responsibility of module owners to decide when it is truly appropriate to have this sort of specialized support, or when the DB api is sufficient (including the potential additional maintenance burden, attractiveness to clients, etc).
On Wed, Feb 21, 2018 at 1:35 AM Daniel-Constantin Mierla miconda@gmail.com wrote:
Hello,
(cross-posting being something I want to get developers and users express their opinions).
We were recently in a (lucky) situation to have two new modules submissions targeting more or less the same purpose: allowing to use Kamailio with a Redis backend via database API.
One was submitted by Andreas Granig (Sipwise) and already merged with the name db_redis, because it was designed from the beginning as a generic DB connector, so the module can be used with auth_db, acc, usrloc, ...
The second one was submitted by Surendra Tiwari (Plivo), initially having a naming conflict with db_redis, but renamed to db_redisusrloc, see the pull request at:
Now, this email is about deciding the way to go forward with the second module.
It was designed to be used only for usrloc in the first phase, with many specific usrloc attributes hard coded inside db_redisusrloc. Surendra said (in a private chat) that the long term plan is to make it work for other modules. Anyhow, at this moment is very tied to usrloc, therefore the name of the module.
Given that the backend (Redis) is not an SQL engine, mapping over Kamailio's DB API needs some schema definition (see the readme of db_redis) in order to make it generic and work for all our modules that use a DB backend.
On the other hand, to squeeze the best of the backend, specially in no-SQL cases, having a dedicated DB connector module optimized for a specific module might help to get more performances/high-availability/scalability from the backend itself.
In this case, for example the expires value for a contact record can be set inside redis, so kamailio module doesn't have to run a timer routine to clean up (it doesn't mean db_redisusrloc does it right now, I didn't have time for a proper review, just giving an example). Surendra said they use it in production for couple of months now and it is several times faster than using usrloc with db_postgres (iirc, not db_mysql) for db_mode=3 (database only mode).
But of course, the reverse of the medal with a dedicated db connector for a module: it adds overhead to code maintenance (besides generic updates due to external library changes, I expect changing something relevant in usrloc, like adding new columns, would require updates in this module as well).
So, there are few things I want feedback on:
- how do you fill about splitting from a generic-only DB connectors to
have also some dedicated ones? This is more from confusion point of view, as a general rule so far, we do not deny contributions if there are other options for same kind of feature (e.g., many lcr or nat traversal options). As long as the contributor is willing to maintain the code, we were fine.
- I guess usrloc, presence and dialog modules would be the main
suspects that would benefit for such dedicated connectors, in other cases might not worth adding dedicated connectors. Missing any other module one would like to squeeze more performances with a dedicated connector?
- should we set a different naming policy for such modules, for
example: use *dbs_* prefix instead *db_*, to suggest better it is a DataBase Specific connector?
- Andreas said he plans to do some performance testing of usrloc module
with the two modules and see the differences. Anyone else that wants to do it? It can be a good metric to see if it worth going one way or another?
- Helping to review the pull request, specially if you use Redis, is
appreciated. Personally I am very short in available time these days, next week I plan do to new Kamailio stable releases, so the schedule is not getting lighter in my side.
Cheers, Daniel
-- Daniel-Constantin Mierla www.twitter.com/miconda -- www.linkedin.com/in/miconda Kamailio Advanced Training - March 5-7, 2018, Berlin - www.asipto.com Kamailio World Conference - May 14-16, 2018 - www.kamailioworld.com
Kamailio (SER) - Users Mailing List sr-users@lists.kamailio.org https://lists.kamailio.org/cgi-bin/mailman/listinfo/sr-users
Hi, just sharing my experience from last year while adding some features to db_postgres and db_sqlite
On Wed, Feb 21, 2018 at 1:33 AM, Daniel-Constantin Mierla <miconda@gmail.com
wrote:
Hello,
(cross-posting being something I want to get developers and users express their opinions).
We were recently in a (lucky) situation to have two new modules submissions targeting more or less the same purpose: allowing to use Kamailio with a Redis backend via database API.
One was submitted by Andreas Granig (Sipwise) and already merged with the name db_redis, because it was designed from the beginning as a generic DB connector, so the module can be used with auth_db, acc, usrloc, ...
The second one was submitted by Surendra Tiwari (Plivo), initially having a naming conflict with db_redis, but renamed to db_redisusrloc, see the pull request at:
Now, this email is about deciding the way to go forward with the second module.
It was designed to be used only for usrloc in the first phase, with many specific usrloc attributes hard coded inside db_redisusrloc. Surendra said (in a private chat) that the long term plan is to make it work for other modules. Anyhow, at this moment is very tied to usrloc, therefore the name of the module.
Given that the backend (Redis) is not an SQL engine, mapping over Kamailio's DB API needs some schema definition (see the readme of db_redis) in order to make it generic and work for all our modules that use a DB backend.
On the other hand, to squeeze the best of the backend, specially in no-SQL cases, having a dedicated DB connector module optimized for a specific module might help to get more performances/high-availability/scalability from the backend itself.
In this case, for example the expires value for a contact record can be set inside redis, so kamailio module doesn't have to run a timer routine to clean up (it doesn't mean db_redisusrloc does it right now, I didn't have time for a proper review, just giving an example). Surendra said they use it in production for couple of months now and it is several times faster than using usrloc with db_postgres (iirc, not db_mysql) for db_mode=3 (database only mode).
But of course, the reverse of the medal with a dedicated db connector for a module: it adds overhead to code maintenance (besides generic updates due to external library changes, I expect changing something relevant in usrloc, like adding new columns, would require updates in this module as well).
So, there are few things I want feedback on:
- how do you fill about splitting from a generic-only DB connectors to
have also some dedicated ones? This is more from confusion point of view, as a general rule so far, we do not deny contributions if there are other options for same kind of feature (e.g., many lcr or nat traversal options). As long as the contributor is willing to maintain the code, we were fine.
Generic is adding a lot of value in reusability but dedicated ones may also be needed, especialy with NoSQL.
Even with SQL DB abstraction we are sometimes facing limitations. Since the db abstraction is generic it may prevent the use of some back-end specific features, even with real SQL back-end when something is *statement specific*, one example : when adding the postgres upsert last year, introduced in Postgres 9.5 the postgres team decided that the conflict/constraints must be specified explicitely in the statement, in this case I decided to do automatic constraint detections so that db_postgres could automaticaly always do upsert when insert_update is used, this was working fine with mysql because insert for update is generic.
In such case may still have several options like extendeding the DB abstraction or bypassing it by presseting specific values/decisions in the db_x module (per connection, table, statement type) but if it is per statement I beleive the only option is to extend the db abstraction.
- I guess usrloc, presence and dialog modules would be the main
suspects that would benefit for such dedicated connectors, in other cases might not worth adding dedicated connectors. Missing any other module one would like to squeeze more performances with a dedicated connector?
I agree, dedicated ones may be valuable, Kamailio is able to do many things but it seems valuable to also be specialised in doing some tasks extremely well.
- should we set a different naming policy for such modules, for
example: use *dbs_* prefix instead *db_*, to suggest better it is a DataBase Specific connector?
- Andreas said he plans to do some performance testing of usrloc module
with the two modules and see the differences. Anyone else that wants to do it? It can be a good metric to see if it worth going one way or another?
- Helping to review the pull request, specially if you use Redis, is
appreciated. Personally I am very short in available time these days, next week I plan do to new Kamailio stable releases, so the schedule is not getting lighter in my side.
Cheers, Daniel
-- Daniel-Constantin Mierla www.twitter.com/miconda -- www.linkedin.com/in/miconda Kamailio Advanced Training - March 5-7, 2018, Berlin - www.asipto.com Kamailio World Conference - May 14-16, 2018 - www.kamailioworld.com
Kamailio (SER) - Users Mailing List sr-users@lists.kamailio.org https://lists.kamailio.org/cgi-bin/mailman/listinfo/sr-users
-
Alex Balashov -
Daniel-Constantin Mierla -
Henning Westerholt -
Josh Converse -
Julien Chavanton