Hi,
As far as I know, "best practice"(in my mind) of lambda mode should looks
like this. Here, I use "batch segment" to refer to segment which source from
Hive, "streaming segment" which source from Kafka.
1. To deal with (most) late income message, you should set
"kylin.stream.cube.duration" to a reasonable value; if 99.5% of would not late
for 2 hour,
you may set kylin.stream.cube.duration to 7200.
2. To deal with some case such as the following scene/case, you may need to
use Kylin's Rest API the build a"batch segment" to replace a "streaming
segment".
1. You want to normalize some value , if some value of "price" is using
dollor($) and others using euro().
2. Correct mistake from kafka message, such as some value are upper
case or other value are capitalize.
3. Some message have been discarded by Streaming Receiver because of
they are beyond the scope of "kylin.stream.cube.duration", but you real need
them.
So, I think build a "batch segment" is only necessary when you find
something wrong and want to overwrite "streaming segment". When you need to
overwrite, you may need to overwrite whole streaming segment, use exact match
segment range, I think in this way segment overlap problem will never happens.
If I misunderstand anything, please let me know. Thank you.
-----------------
-----------------
Best wishes to you !
From ：Xiaoxiang Yu
At 2019-06-26 19:36:44, "Andras Nagy" <[email protected]> wrote:
Hi Xiaoxiang, ShaoFeng,
Thank you for your answers!
Regarding the segment overlap between batch and streaming, my point was that it
seems to be different to how I understand segment overlap to work in streaming
OLAP.
That is, assuming I build a "batch" segment from 2019-06-25 00:00:00.0 to
2019-06-26 00:00:00.0 (1 day).
Then if a late event comes in for the same period (e.g. event timestamp field
contains 2019-06-25 12:34:56), but after this batch segment has already been
built, it will not show up in the query result unless I set up a mechanism to
detect the late event and trigger the rebuilding of the batch segment. This is
because the results from the batch segments overwrite the results from the
streaming segments.
On the other hand, for segments built by the streaming engine, my understanding
is that they can have overlapping time periods and the query engine will merge
the results.
I understand this behaviour is actually useful in optimizing the query path in
case there were many overlapping segments created by the streaming cube build,
since with the batch-built segment, the results can be served from a single
segment and don't need to be merged from multiple overlapping segments.
I guess the solution here is to ensure that the batch segment is always built
for a time period from which we practically don't expect late events anymore.
Best regards,
Andras
On Wed, Jun 26, 2019 at 11:40 AM Xiaoxiang Yu <[email protected]> wrote:
Hi Andras, Shaofeng,
I will update this information asap.
About segment overlaping problem, I have a test in my env, looks like
everything works well. Since the segment range created by kylin’s streaming
coordinator is something like "201906290000_201906290100" , if you want to
build a segment, I think you should use the exact match segment range (such as
"201906290000_201906290100"), or merge multi exist segments range (such as
"201906290100_201906290300") .
-----------------
-----------------
Best wishes to you !
From ：Xiaoxiang Yu
At 2019-06-26 12:00:38, "ShaoFeng Shi" <[email protected]> wrote:
Hi Xiaoxiang,
Thank you for the detailed information. Could you please record these
limitations as JIRA issues (if not yet)? Thanks.
Best regards,
Shaofeng Shi 史少锋
Apache Kylin PMC
Email: [email protected]
Apache Kylin FAQ: https://kylin.apache.org/docs/gettingstarted/faq.html
Join Kylin user mail group: [email protected]
Join Kylin dev mail group: [email protected]
Xiaoxiang Yu <[email protected]> 于2019年6月25日周二 下午11:42写道：
Hi, Andras
I am glad to see that you have have a strong understanding with Kylin's
Realtime OLAP. Most of them are correct, the following is my understanding:
1) Currently, there is no such documentation which talk about how to use
lambda mode, we will publish one after 3.0.0-beta release (maybe this wekend or
after a week?).
2) Hive table must have the same name as the streaming table , and should
be locate at "default" namespace of hive. The column name should match exactly
and data type should be compatible.
3) If you want to build segment which data from hive, you have to built
by rest api.
4) Cube build engine must be mapreduce, spark is not supported at the
moment.
-----------------
-----------------
Best wishes to you !
From ：Xiaoxiang Yu
At 2019-06-25 17:20:55, "Andras Nagy" <[email protected]> wrote:
Hi ShaoFeng,
Thanks a lot for the pointer on the lambda mode, yes, that's exactly what I
need :)
Is there perhaps documentation on this? For now, I was trying to get this
working 'empirically' and finally succeeded, but some of my conclusions may be
wrong. This is what I concluded:
- hive table must have the same name as the streaming table (name given to the
data source)
- cube can't be built from UI (to build the historic segments from the data in
hive), but it can be built using the REST API
- cube build engine must be mapreduce. For Spark as build engine I got
exception "Cannot adapt to interface org.apache.kylin.engine.spark.ISparkOutput"
- endTime must be non-overlapping with the streaming data. When I had overlap,
the streaming data coming from kafka did not show up in the output, I guess
this is what you meant by "the segments from Hive will overwrite the segments
from Kafka".
Are these correct conclusions? Is there anything else I should be aware of?
Many thanks,
Andras
On Tue, Jun 25, 2019 at 9:19 AM ShaoFeng Shi <[email protected]> wrote:
Hello Andras,
Kylin's realtime-OLAP feature supports a "Lambda" mode (mentioned in
https://kylin.apache.org/blog/2019/04/12/rt-streaming-design/), which means,
you can define a fact table whose data can be from both Kafka and Hive. The
only requirement is that all the cube columns appear in both Kafka data and
Hive data. I think maybe that can fit your need. The cube can be built from
Kafka, in the meanwhile, it can also be built from Hive, the segments from Hive
will overwrite the segments from Kafka (as usually Hive data is more accurate).
When querying the cube, Kylin will firstly query historical segments, and then
real-time segments (adding the max-time of historical segments as the
condition).
Best regards,
Shaofeng Shi 史少锋
Apache Kylin PMC
Email: [email protected]
Apache Kylin FAQ: https://kylin.apache.org/docs/gettingstarted/faq.html
Join Kylin user mail group: [email protected]
Join Kylin dev mail group: [email protected]
Andras Nagy <[email protected]> 于2019年6月24日周一 下午11:29写道：
Dear Ma,
Thanks for your reply.
Slightly related to my original question on the hybrid model, I was wondering
if it's possible to combine a batch and a streaming cube. I realized this is
not possible, as a hybrid model can only be created from cubes of the same
model (and a model points to either a batch or a streaming datasource).
The usecase would be this:
- we have a large amount of streaming data in Kafka that we would like to
process with Kylin streaming
- Kafka retention is only a few days, so if we need to change anything in the
cubes (e.g. introduce a new metric or dimension which has been present in the
events, but not in the cube definition), we can only reprocess a few days worth
of data in the streaming model
- the raw events are also written to a data lake for long-term storage
- the data written to the data lake could be used to feed the historic data
into a batch kylin model (and cubes)
- I'm looking for a way to combine these, so if we want to change anything in
the cubes, we can recalculate them for the historic data as well
Is there a way to achieve this with current Kylin? (Without implementing a
custom query layer that combines the two cubes.)
Best regards,
Andras
On Fri, Jun 14, 2019 at 6:43 AM Ma Gang <[email protected]> wrote:
Hi Andras,
Currently it doesn't support consume from specified offsets, only support
consume from startOffset or latestOffset, if you want to consume from
startOffset, you need to set the configuration:
kylin.stream.consume.offsets.latest to false in the cube's overrides page.
If you do need to start from specified offsets, please create a jira request,
but I think it is hard for user to know what's the offsets should be set for
all partitions.
At 2019-06-13 22:34:59, "Andras Nagy" <[email protected]> wrote:
Dear Ma,
Thank you very much!
>1)yes, you can specify a configuration in the new cube, to consume data from
>start offset
That is, an offset value for each partition of the topic? That would be good -
could you please point me where to do this in practice, or point me to what I
should read? (I haven't found it on the cube designer UI - perhaps this is
something that's only available on the API?)
Many thanks,
Andras
On Thu, Jun 13, 2019 at 1:14 PM Ma Gang <[email protected]> wrote:
Hi Andras,
1)yes, you can specify a configuration in the new cube, to consume data from
start offset
2)It should work, but I haven't tested it yet
3)as I remember, currently we use Kafka 1.0 client library, so it is better to
use the version later, I'm sure that the version before 0.9.0 cannot work, but
not sure 0.9.x can work or not
| |
Ma Gang
|
|
邮箱：[email protected]
|
签名由 网易邮箱大师 定制
On 06/13/2019 18:01, Andras Nagy wrote:
Greetings,
I have a few questions related to the new streaming (real-time OLAP)
implementation.
1) Is there a way to have data reprocessed from kafka? E.g. I change a cube
definition and drop the cube (or add a new cube definition) and want to have
data that is still available on kafka to be reprocessed to build the changed
cube (or new cube)? Is this possible?
2) Does the hybrid model work with streaming cubes (to combine two cubes)?
3) What is minimum kafka version required? The tutorial asks to install Kafka
1.0, is this the minimum required version?
Thank you very much,
Andras
