Amazon Kinesis Data Analytics
Developer Guide

Timestamps and the ROWTIME Column

In-application streams include a special column called ROWTIME. It stores a timestamp when Amazon Kinesis Data Analytics inserts a row in the first in-application stream. ROWTIME reflects the timestamp at which Amazon Kinesis Data Analytics inserted a record into the first in-application stream after reading from the streaming source. This ROWTIME value is then maintained throughout your application.

Note

When you pump records from one in-application stream into another, you don't need to explicitly copy the ROWTIME column, Amazon Kinesis Data Analytics copies this column for you.

Amazon Kinesis Data Analytics guarantees that the ROWTIME values are monotonically increased. You use this timestamp in time-based windowed queries. For more information, see Windowed Queries.

You can access the ROWTIME column in your SELECT statement like any other columns in your in-application stream. For example:

SELECT STREAM ROWTIME, some_col_1, some_col_2 FROM SOURCE_SQL_STREAM_001

Understanding Various Times in Streaming Analytics

In addition to ROWTIME, there are other types of times in real-time streaming applications. These are:

  • Event time – The timestamp when the event occurred. This is also sometimes called the client-side time. It is often desirable to use this time in analytics because it is the time when an event occurred. However, many event sources, such as mobile phones and web clients, do not have reliable clocks, which can lead to inaccurate times. In addition, connectivity issues can lead to records appearing on a stream not in the same order the events occurred.

     

  • Ingest time – The timestamp of when record was added to the streaming source. Amazon Kinesis Data Streams includes a field called APPROXIMATE_ARRIVAL_TIME in every record that provides this timestamp. This is also sometimes referred to as the server-side time. This ingest time is often the close approximation of event time. If there is any kind of delay in the record ingestion to the stream, this can lead to inaccuracies, which are typically rare. Also, the ingest time is rarely out of order, but it can occur due to the distributed nature of streaming data. Therefore, Ingest time is a mostly accurate and in-order reflection of the event time.

     

  • Processing time – The timestamp when Amazon Kinesis Data Analytics inserts a row in the first in-application stream. Amazon Kinesis Data Analytics provides this timestamp in the ROWTIME column that exists in each in-application stream. The processing time is always monotonically increasing, but it will not be accurate if your application falls behind (if an application falls behind, the processing time will not accurately reflect the event time). This ROWTIME is very accurate in relation to the wall clock, but it might not be the time when the event actually occurred.

As you can see from the preceding discussion, using each of these times in windowed queries that are time-based has advantages and disadvantages. We recommend you choose one or more of these times, and a strategy to deal with the relevant disadvantages based on your use case scenario.

Note

If you are using row-based windows, time is not an issue and you can ignore this section.

We recommend a two-window strategy that uses two time-based, both ROWTIME and one of the other times (ingest or event time).

  • Use ROWTIME as the first window, which controls how frequently the query emits the results, as shown in the following example. It is not used as a logical time.

  • Use one of the other times that is the logical time you want to associated with your analytics. This time represents when the event occurred. In the following example, the analytics goal is to group the records and return count by ticker.

The advantage of this strategy is that it can use a time that represents when the event occurred, and it can gracefully handle when your application falls behind or when events arrive out of order. If the application falls behind when bringing records into the in-application stream, they are still grouped by the logical time in the second window. The query uses ROWTIME to guarantee the order of processing. Any records that are late (ingest timestamp shows earlier value compared to the ROWTIME value) are processed successfully too.

Consider the following query against the demo stream used in the Getting Started Exercise. The query uses the GROUP BY clause and emits ticker count in a one-minute tumbling window.

CREATE OR REPLACE STREAM "DESTINATION_SQL_STREAM" ("ingest_time" timestamp, "APPROXIMATE_ARRIVAL_TIME" timestamp, "ticker_symbol" VARCHAR(12), "symbol_count" integer); CREATE OR REPLACE PUMP "STREAM_PUMP" AS INSERT INTO "DESTINATION_SQL_STREAM" SELECT STREAM STEP("SOURCE_SQL_STREAM_001".ROWTIME BY INTERVAL '60' SECOND) AS "ingest_time", STEP("SOURCE_SQL_STREAM_001".APPROXIMATE_ARRIVAL_TIME BY INTERVAL '60' SECOND) AS "APPROXIMATE_ARRIVAL_TIME", "TICKER_SYMBOL", COUNT(*) AS "symbol_count" FROM "SOURCE_SQL_STREAM_001" GROUP BY "TICKER_SYMBOL", STEP("SOURCE_SQL_STREAM_001".ROWTIME BY INTERVAL '60' SECOND), STEP("SOURCE_SQL_STREAM_001".APPROXIMATE_ARRIVAL_TIME BY INTERVAL '60' SECOND);

In GROUP BY, you first group the records based on ROWTIME in a one-minute window and then by APPROXIMATE_ARRIVAL_TIME.

Note that the timestamp values in the result are rounded down to the nearest 60 second interval. The first group result emitted by the query shows records in the first minute. The second group of results emitted shows records in the next minutes based on ROWTIME. The last record indicates that the application was late in bringing the record in the in-application stream (it shows a late ROWTIME value compared to the ingest timestamp).

ROWTIME INGEST_TIME TICKER_SYMBOL SYMBOL_COUNT --First one minute window. 2016-07-19 17:05:00.0 2016-07-19 17:05:00.0 ABC 10 2016-07-19 17:05:00.0 2016-07-19 17:05:00.0 DEF 15 2016-07-19 17:05:00.0 2016-07-19 17:05:00.0 XYZ 6 –-Second one minute window. 2016-07-19 17:06:00.0 2016-07-19 17:06:00.0 ABC 11 2016-07-19 17:06:00.0 2016-07-19 17:06:00.0 DEF 11 2016-07-19 17:06:00.0 2016-07-19 17:05:00.0 XYZ 1 *** ***late-arriving record, instead of appearing in the result of the first 1-minute windows (based on ingest_time, it is in the result of the second 1-minute window.

You can combine the results for a final accurate count per minute by pushing the results to a downstream database. For example, you can configure application output to persist the results to a Kinesis Data Firehose delivery stream that can write to an Amazon Redshift table. After results are in an Amazon Redshift table, you can query the Amazon Redshift table to compute the total count group by Ticker_Symbol. In the case of XYZ, the total is accurate (6+1) even though a record arrived late.