Обсуждение: Load a csv or a avro?
Hello all,
Its postgres database. We have option of getting files in csv and/or in avro format messages from another system to load it into our postgres database. The volume will be 300million messages per day across many files in batches.
My question was, which format should we chose in regards to faster data loading performance ? and if any other aspects to it also should be considered apart from just loading performance?
Hi
There are different data formats available, following are few points for there performance implications
1. CSV : It's easy to use and widely supported but it can be slower due to parsing overload.
2. Binary : Its faster to load but not human understandable.
Hope this helps.
Regards
Kashif Zeeshan
On Fri, Jul 5, 2024 at 2:08 PM sud <suds1434@gmail.com> wrote:
Hello all,
Its postgres database. We have option of getting files in csv and/or in avro format messages from another system to load it into our postgres database. The volume will be 300million messages per day across many files in batches.
My question was, which format should we chose in regards to faster data loading performance ? and if any other aspects to it also should be considered apart from just loading performance?
pá 5. 7. 2024 v 11:08 odesílatel sud <suds1434@gmail.com> napsal: > > Hello all, > > Its postgres database. We have option of getting files in csv and/or in avro format messages from another system to loadit into our postgres database. The volume will be 300million messages per day across many files in batches. > > My question was, which format should we chose in regards to faster data loading performance ? and if any other aspectsto it also should be considered apart from just loading performance? We are able to load ~300 million rows per one day using CSV and COPY functions (https://www.postgresql.org/docs/current/libpq-copy.html#LIBPQ-COPY-SEND).
Hi,
Performance Considerations
Avro files are smaller due to compression so needing less I/O time. whereas CSV files are simpler but larger in size so read/write will need more time.
COPY command works very well with CSV files whereas ETL process is required for handling Avro.
Regards,
Muhammad Ikram
On Fri, Jul 5, 2024 at 3:03 PM Josef Šimánek <josef.simanek@gmail.com> wrote:
pá 5. 7. 2024 v 11:08 odesílatel sud <suds1434@gmail.com> napsal:
>
> Hello all,
>
> Its postgres database. We have option of getting files in csv and/or in avro format messages from another system to load it into our postgres database. The volume will be 300million messages per day across many files in batches.
>
> My question was, which format should we chose in regards to faster data loading performance ? and if any other aspects to it also should be considered apart from just loading performance?
We are able to load ~300 million rows per one day using CSV and COPY
functions (https://www.postgresql.org/docs/current/libpq-copy.html#LIBPQ-COPY-SEND).
Muhammad Ikram
On Fri, Jul 5, 2024 at 5:08 AM sud <suds1434@gmail.com> wrote:
Hello all,
Its postgres database. We have option of getting files in csv and/or in avro format messages from another system to load it into our postgres database. The volume will be 300million messages per day across many files in batches.
My question was, which format should we chose in regards to faster data loading performance ?
What application will be loading the data? If psql, then go with CSV; COPY is really efficient.
If the PG tables are already mapped to the avro format, then maybe avro will be faster.
and if any other aspects to it also should be considered apart from just loading performance?
If all the data comes in at night, drop as many indices as possible before loading.
Load each file in as few DB connections as possible: the most efficient binary format won't do you any good if you open and close a connection for each and every row.
On 7/5/24 02:08, sud wrote: > Hello all, > > Its postgres database. We have option of getting files in csv and/or in > avro format messages from another system to load it into our postgres > database. The volume will be 300million messages per day across many > files in batches. Are dumping the entire contents of each file or are you pulling a portion of the data out? > > My question was, which format should we chose in regards to faster data > loading performance ? and if any other aspects to it also should be > considered apart from just loading performance? > -- Adrian Klaver adrian.klaver@aklaver.com
On Fri, Jul 5, 2024 at 3:27 PM Kashif Zeeshan <kashi.zeeshan@gmail.com> wrote:
HiThere are different data formats available, following are few points for there performance implications1. CSV : It's easy to use and widely supported but it can be slower due to parsing overload.2. Binary : Its faster to load but not human understandable.Hope this helps.RegardsKashif Zeeshan
My understanding was that it will be faster to load .csv as it is already being mapped to table rows and columns whereas in case of .avro the mapping has to be done so that the fields in the avro can be mapped to the columns in the table appropriately and that will be having additional overhead. Is my understanding correct?
On Fri, Jul 5, 2024 at 8:24 PM Adrian Klaver <adrian.klaver@aklaver.com> wrote:
On 7/5/24 02:08, sud wrote:
> Hello all,
>
> Its postgres database. We have option of getting files in csv and/or in
> avro format messages from another system to load it into our postgres
> database. The volume will be 300million messages per day across many
> files in batches.
Are dumping the entire contents of each file or are you pulling a
portion of the data out?
Yes, all the fields in the file have to be loaded to the columns in the tables in postgres. But how will that matter here for deciding if we should ask the data in .csv or .avro format from the outside system to load into the postgres database in row and column format? Again my understanding was that irrespective of anything , the .csv file load will always faster as because the data is already stored in row and column format as compared to the .avro file in which the parser has to perform additional job to make it row and column format or map it to the columns of the database table. Is my understanding correct here?
On Sat, Jul 6, 2024 at 4:10 PM sud <suds1434@gmail.com> wrote:
On Fri, Jul 5, 2024 at 8:24 PM Adrian Klaver <adrian.klaver@aklaver.com> wrote:On 7/5/24 02:08, sud wrote:
> Hello all,
>
> Its postgres database. We have option of getting files in csv and/or in
> avro format messages from another system to load it into our postgres
> database. The volume will be 300million messages per day across many
> files in batches.
Are dumping the entire contents of each file or are you pulling a
portion of the data out?Yes, all the fields in the file have to be loaded to the columns in the tables in postgres.
But you didn't say which columns or which tables.
If one row of CSV input must be split into multiple tables, then it might be pretty slow.
But how will that matter here for deciding if we should ask the data in .csv or .avro format from the outside system to load into the postgres database in row and column format? Again my understanding was that irrespective of anything , the .csv file load will always faster as because the data is already stored in row and column format as compared to the .avro file in which the parser has to perform additional job to make it row and column format or map it to the columns of the database table. Is my understanding correct here?
Yes and no. It all depends on how well each input row maps to a Postgresql table.
Bottom line: you want an absolute answer, but we can't give you an absolute answer, since we don't know what the input data looks like, and we don't know what the Postgresql tables look like.
An AVRO file might be faster to input than CSV, or it might be horribly slower.
And you might incompetently program a CSV importer so that it's horribly slow.
We can't give absolute answers without knowing more details than the ambiguous generalities in your emails.
On 7/6/24 13:09, sud wrote: > On Fri, Jul 5, 2024 at 8:24 PM Adrian Klaver <adrian.klaver@aklaver.com > <mailto:adrian.klaver@aklaver.com>> wrote: > > On 7/5/24 02:08, sud wrote: > > Hello all, > > > > Its postgres database. We have option of getting files in csv > and/or in > > avro format messages from another system to load it into our > postgres > > database. The volume will be 300million messages per day across many > > files in batches. > > Are dumping the entire contents of each file or are you pulling a > portion of the data out? > > > > Yes, all the fields in the file have to be loaded to the columns in the > tables in postgres. But how will that matter here for deciding if we > should ask the data in .csv or .avro format from the outside system to > load into the postgres database in row and column format? Again my > understanding was that irrespective of anything , the .csv file load > will always faster as because the data is already stored in row and > column format as compared to the .avro file in which the parser has to > perform additional job to make it row and column format or map it to the > columns of the database table. Is my understanding correct here? If you are going to use complete rows and all rows then COPY of CSV in Postgres would be your best choice. -- Adrian Klaver adrian.klaver@aklaver.com