This describes the analysis of the legacy-data-import which took way too long, which turned out to be a problem in the RBAC-access-rights-check as well as `EntityManager.persist` creating too many SQL queries.
During the legacy data import for hosting assets we noticed massive performance problems. The import of about 2200 hosting-assets (IP-numbers, managed-webspaces, managed- and cloud-servers) as well as the creation of booking-items and booking-projects as well as necessary office-data entities (persons, contacts, partners, debitors, relations) **took 25 minutes**.
Importing hosting assets up to UnixUsers and EmailAddresses even **took about 100 minutes**.
(The office data import sometimes, but rarely, took only 10min.
We could not find a pattern, why that was the case. The impression that it had to do with too many other parallel processes, e.g. browser with BBB or IntelliJ IDEA was proved wrong, but stopping all unnecessary processes and performing the import again.)
That both queries total up to more than the runtime of the import-process is most likely due to internal parallel query processing.
## Attempts to Mitigate the Problem
### VACUUM ANALYZE
In the middle of the import, we updated the PostgreSQL statistics to recalibrate the query optimizer:
```SQL
VACUUM ANALYZE;
```
This did not improve the performance.
### Improving Joins + Indexes
We were suspicious about the sequential scan over all `rbacpermission` rows which was done by PostgreSQL to execute a HashJoin strategy. Turning off that strategy by
-> CTE Scan on grants (cost=0.00..11.84 rows=592 width=16)
[...]
```
The HashJoin strategy could be great if the hash-map could be kept for multiple invocations. But during an import process, of course, there are always new rows in the underlying table and the hash-map would be outdated immediately.
Also creating indexes which should suppor the RBAC query, like the following, did not improve performance:
```SQL
create index on RbacPermission (objectUuid, op);
create index on RbacPermission (opTableName, op);
```
### LAZY loading for Relation.anchorPerson/.holderPerson/
At this point, the import took 21mins with these statistics:
| select hore1_0.uuid,a1_0.uuid,a1_0.familyname,a1_0.givenname,a1_0.persontype,a1_0.salutation,a1_0.title,a1_0.tradename,a1_0.version,c1_0.uuid,c1_0.caption,c1_0.emailaddresses,c1_0.phonenumbers,c1_0.postaladdress, c1_0.version,h1_0.uuid,h1_0.familyname,h1_0.givenname,h1_0.persontype,h1_0.salutation,h1_0.title,h1_0.tradename,h1_0.version,hore1_0.mark,hore1_0.type,hore1_0.version from public.hs_office.relation_rv hore1_0 left join public.hs_office.person_rv a1_0 on a1_0.uuid=hore1_0.anchoruuid left join public.hs_office.contact_rv c1_0 on c1_0.uuid=hore1_0.contactuuid left join public.hs_office.person_rv h1_0 on h1_0.uuid=hore1_0.holderuuid where hore1_0.uuid=$1 | 517 | 11 | 1282 |
| select hope1_0.uuid,hope1_0.familyname,hope1_0.givenname,hope1_0.persontype,hope1_0.salutation,hope1_0.title,hope1_0.tradename,hope1_0.version from public.hs_office.person_rv hope1_0 where hope1_0.uuid=$1 | 973 | 4 | 254 |
| select hoce1_0.uuid,hoce1_0.caption,hoce1_0.emailaddresses,hoce1_0.phonenumbers,hoce1_0.postaladdress,hoce1_0.version from public.hs_office.contact_rv hoce1_0 where hoce1_0.uuid=$1 | 973 | 4 | 253 |
| with recursive grants as ( select descendantUuid, ascendantUuid from RbacGrants where descendantUuid = grantedId union all select ""grant"".descendantUuid, ""grant"".ascendantUuid from RbacGrants ""grant"" inner join grants recur on recur.ascendantUuid = ""grant"".descendantUuid ) select exists ( select $3 from grants where ascendantUuid = any(granteeIds) ) or grantedId = any(granteeIds) | 47540 | 0 | 0 |
| insert into RbacGrants (grantedByTriggerOf, ascendantuuid, descendantUuid, assumed) values (currentTriggerObjectUuid(), superRoleId, subRoleId, doAssume) on conflict do nothing" | 40472 | 0 | 0 |
The slowest query now was fetching Relations joined with Contact, Anchor-Person and Holder-Person, for all tables using the restricted (RBAC) views (_rv).
We changed these mappings from `EAGER` (default) to `LAZY` to `@ManyToOne(fetch = FetchType.LAZY)` and got this result:
| with recursive grants as ( select descendantUuid, ascendantUuid from RbacGrants where descendantUuid = grantedId union all select ""grant"".descendantUuid, ""grant"".ascendantUuid from RbacGrants ""grant"" inner join grants recur on recur.ascendantUuid = ""grant"".descendantUuid ) select exists ( select $3 from grants where ascendantUuid = any(granteeIds) ) or grantedId = any(granteeIds) | 47538 | 0 | 0 |
insert into RbacGrants (grantedByTriggerOf, ascendantuuid, descendantUuid, assumed) values (currentTriggerObjectUuid(), superRoleId, subRoleId, doAssume) on conflict do nothing | 40472 | 0 | 0 |
Now, finally, the total runtime of the import was down to 12 minutes. This is repeatable, where originally, the import took about 25mins in most cases and just rarely - and for unknown reasons - 10min.
But once UnixUser and EmailAlias assets got added to the import, the total time went up to about 110min.
This was not acceptable, especially not, considering that domains, email-addresses and database-assets are almost 10 times that number and thus the import would go up to over 1100min which is 20 hours.
In a first step, a `HsHostingAssetRawEntity` was created, mapped to the raw table (hs_hosting.asset) not to the RBAC-view (hs_hosting.asset_rv). Unfortunately we did not keep measurements, but that was only part of the problem anyway.
The main problem was, that there is something strange with persisting (`EntityManager.persist`) for EmailAlias assets. Where importing UnixUsers was mostly slow due to RBAC SELECT-permission checks, persisting EmailAliases suddenly created about a million (in numbers 1.000.000) SQL UPDATE statements after the INSERT, all with the same data, just increased version number (used for optimistic locking). We were not able to figure out why this happened.
Keep in mind, it's the same table with the same RBAC-triggers, just a different value in the type column.
Once `EntityManager.persist` was replaced by an explicit SQL INSERT - just for `HsHostingAssetRawEntity`, the total time was down to 17min. Thus importing the UnixUsers and EmailAliases took just 5min, which is an acceptable result. The total import of all HostingAssets is now estimated to about 1 hour (on my developer laptop).
It seems that the trigger effects (eg. No. 3 and No. 4) are included in the measure for the causing INSERT, otherwise summing up the totals would exceed the actual total time of the whole import. And it was to be expected that building the RBAC rules for new business objects takes most of the time.
In production, the `SELECT ... FROM hs_office.relation_rv` (No. 2) with about 0.5 seconds could still be a problem. But once we apply the improvements from the hosting asset area also to the office area, this should not be a problem for the import anymore.
The origin problem was the expensive RBAC check for many SELECT queries.
This consists of two parts:
1. The recursive CTE query to determine which object's UUIDs are visible for the current subject.
This query itself takes currently about 250ms thus is no problem by itself as long as we only need it once per request.
2. Joining the result from 1. with the result if a business query.
The performance of the business query itself is no problem, for the join see the following explanations.
Superusers can see all objects (currently already over 90.000)
and even high level roles of customers with many hosting assets can see several thousand objects.
This is the one side of that problematic join.
The other side of that problematic is the result of the business query.
For example if a user wants to select all of their e-mail-addresses, that might easily half of the visible objects.
Thus, we would have a join of for example 5.000 x 2.500 rows, which is going to be slow.
As there are currently about 84.000 objects are hosting assets and 33.000 e-mail-addresses in our system,
for a superuser we would even run into an 84.0000 x 33.0000 join.
We found some solution approaches:
1. Getting rid of the `rbacrole` and `rbacpermission` table and only having implicit roles with implicit grants (OWNER->ADMIN->AGENT->TENENT->REFERRER) by comparison of ordered enum values and fixed permission assignments (e.g. OWENER->DELETE, ADMIN->UPDATE etc.). We could also get rid of the table `rbacreferece` if we enter users as business objects.
This should dramatically reduce the size of the table `rbackgrant` as well as the recusion levels.
But since we only apply this query once for each business query, that would only improve performance once we have way more objects in our system, but does not help our current problem.
It's quite some effort to implement even just a prototype, so we did not further explore this idea.
2. Adding the object type to the table `rbacObject` to reduce the size of the result of the recursive CTE query.
See chapter below.
3. Inverting the recursion of the CTE-query, combined with the type condition.
To evaluate this, I added a type column to the `rbacObject` table, initially as an enum hsHostingAssetType. Then I entered the type there for all rows from hs_hosting.asset. This means that 83,886 of 92,545 rows in `rbacobject` have a type set, leaving 8,659 without.
If we do this for other types (we currently have 1,271 relations and 927 booking items), it gets more complicated because they are different enum types. As varchar(16), we could lose performance again due to the higher storage space requirements.
But the performance gained is not particularly high anyway.
| currently (without type comparision in rbacobject): | ~3.30 - ~3.49 | ~0.23 |
| optimized (with type comparision in rbacobject): | ~2.99 - ~3.08 | ~0.21 |
As you can see, the query is no problem at all for normal customers (in the example, yours truly). With Hostsharing (D-1000000-hsh) it is quite slow.
Luckily this experiment also shows that it's not a big problem, having all hosting assets in the same database table.
Implementing this approach would be a bit difficult anyway, because we would need to transfer the type query parameter into the definition of the restricted view. We have not even the slightest idea how this could be done.
See the related queries in [recursive-cte-experiments-for-accessible-uuids.sql](../sql/recursive-cte-experiments-for-accessible-uuids.sql). They might have changed independently since this document was written, but you can still check out the old version from git.
### Rearranging the Parts of the CTE-Query
I also moved the function call which determines into its own WITH-section, with no improvement.
Experimentally I moved the business condition into the CTE SELECT, also with no improvement.
Such rearrangements seem to be successfully done by the PostgreSQL query optimizer.
Rearranging the CTE query by extracting parts into WITH-clauses did not improve the performance.
Surprisingly little performance gain (<10%improvement)camefromreducingtheresultoftheCTEquerybymovingthehostingassettypeintoRBAC-systemandusingitintheinnerSELECTqueryinsteadofintheouterSELECTqueryoftheapplicationside.
The major improvement came from using direct INSERT statements, which avoided some SELECT statements unnecessarily generated by the EntityManager and also completely bypassed the RBAC SELECT permission checks.
Where this performance analysis was mostly helping the performance of the legacy data import, we still need measures and improvements for the productive code.
For sure, using more LAZY-loading also helps in the production code. For some more ideas see section _Further Options To Explore_.
