snowflake-cloud-data-platformsnowflake-pipe

How to Move or Alter a Pipe without missing or duplicating any records


This page on managing pipes suggests a process for altering the copy into statement in a pipe.

  1. Pause the pipe (using ALTER PIPE … SET PIPE_EXECUTION_PAUSED=true).
  2. Query the SYSTEM$PIPE_STATUS function and verify that the pipe execution state is PAUSED and the pending file count is 0.
  3. Recreate the pipe to change the COPY statement in the definition. Choose either of the following options: Drop the pipe (using DROP PIPE) and create it (using CREATE PIPE). Recreate the pipe (using the CREATE OR REPLACE PIPE syntax). Internally, the pipe is dropped and created.
  4. Query the SYSTEM$PIPE_STATUS function again and verify that the pipe execution state is RUNNING.

However, if a file should be loaded during the time between pausing and recreating the pipe, there's no step here to refresh for that gap. Even if these steps happen quickly, we've had examples of files getting missed.

Running an ALTER PIPE REFRESH though, causes duplicates because copy history is tied to the pipe. The recreated pipe doesn't have this history and will go back and reload everything.

Is there a nice way to script a change like this out to guarantee that there's no gaps or overlap? Something like getting the timestamp for when the original pipe was paused, and then using that timestamp in a refresh query?

Update: We built a full process and combination of scripts to handle our scenarios. Full script Included in answer below.


Solution

  • We wrote up this process/script for managing snowpipes. It includes validating and refreshing files that were dropped when the pipe was altered or moved.

    It is somewhat specific to our processes and standards, but it should be easily modified/generalized depending on use case.

    /* =====================================================================
    This script provides tools for managing, altering, migrating snowpipes
        and their underlying tables. These instructions may need to be tweaked
        depending on your exact use case.
    
    Assuming the pipes and stages follow our standard naming conventions,
        you can find and replace <Database_Name>, <Schema_Name>, <Table_Name>
        with their respective values
    
    ======================================================================== */
    
    ------------------------------------------
    -- Set up Context and Variables
    ------------------------------------------
    --Set your context so you don’t accidently run scripts in the wrong place
    use <Database_Name>.<Schema_Name>
    
    --Pause the pipe, also giving us the pipe altered time
    --Prefer to avoid running this at the top of the hour to make it easier to verify file loads below.
    alter pipe <Table_Name>_PIPE set pipe_execution_paused = true;
    
    --Get timestamp for when pipe was paused and set as variable $pipe_altered_time
    set pipe_altered_time = (select last_altered::timestamp_tz from information_schema.pipes where pipe_name = '<Table_Name>_PIPE');
    
    --View variable from above to verify that it was altered just now
    select $pipe_altered_time;
    
    ------------------------------------------
    -- Alter table as per request
    ------------------------------------------
    --This step will change depending on the resquest. E.g. it can be skipped if you are just moving a pipe
    
        --Example for adding columns to a pipe
        alter table <Table_Name> add column <CREATE_DATETIME_UTC> <timestamp_NTZ>;
        alter table <Table_Name> add column <INCOMING_RAW_REQUEST_ID> <STRING>;
        alter table <Table_Name> add column <INITIAL_PING_ID> <INT>;
        --Viewing if new column was added
        --Note: Make sure the order of all of the columns matches the pipe column order below
        select top 10 * from <Table_Name>;
    
        --Example for migrating a pipe
        create table <New_Table_Name> clone <Old_Database_Name>.<Old_Schema_Name>.<Old_Table_Name>
    
    ------------------------------------------
    -- Alter Pipe
    ------------------------------------------
    --Altered pipe statement and turns the pipe on
    --This DDL should be retrieved using `SELECT GET_DDL('pipe', '<Table_Name>_PIPE')` (in the same schema context) and then modified to add/adjust columns if needed.
    --Make sure that extracted column names here come from the JSON attribute names, which should be provided in the request.
    --Most Pipes should follow a very similar pattern to the example below
    create or replace pipe <Table_Name>_PIPE
        auto_ingest=true 
        integration='EVENT_HUB_QUEUE' 
        as 
    <
       COPY INTO <Table_Name>
        FROM (SELECT PARSE_JSON(HEX_DECODE_STRING($1:Body)) as JSON, 1 as IS_ACTIVE, metadata$filename, JSON:CreatedDateUtc, JSON:Id, JSON:InitialPingId FROM @<Table_Name>_STAGE)
    >;
    
    --"Refresh" pipe. Reprocess files that were loaded during the paused time frame. 
    -- This can take a few minutes depending on the table and how long it was paused.
    alter pipe <Table_Name>_PIPE REFRESH MODIFIED_AFTER = $pipe_altered_time;
    
    -------------------------------------------------------------------------------------------------------
    --VALIDATION
    -------------------------------------------------------------------------------------------------------
    
    ------------------------------------------
    -- Verify that new files are being processed as expected
    ------------------------------------------
    
    --Data validation - see what files have been loaded
    --It can take a while for new files to show up depending on the event_hub.
    --The real thing you want to check is that the pipe is working, and you see files from after the time when the pipe was paused
    --If there are new files coming in from Event_hub, but not making it to the table, the pipe may have errors that will help for troubleshooting
    select LAST_LOAD_TIME, ROW_COUNT, FILE_NAME
      from table(information_schema.copy_history(Table_Name=>'<Table_Name>', start_time=> dateadd(hours, -2, current_timestamp())))
      ORDER BY LAST_LOAD_TIME DESC;
    
    --Validate data made it to the new columns
    --Make sure that there are new records since the $pipe_altered_time in the table (it might take a while for them to show up)
    --  Validate that the new records look as expected
    --  If there is confusion here, it would be appropriate to contact the requestor on if the data looks correct
    select * from <Table_Name>
    where <any new column name> is not null;
    
    ------------------------------------------
    --compare files in the storage account vs files loaded into the table
    ------------------------------------------
    
    --Choose an hour of the day for comparing (Generally the hour during which the pipe was paused ($pipe_altered_time))
    --see what files have been loaded recently if needed
    select distinct top 1000 file_name from <Table_Name>
    order by file_name desc;
    
    --run these to create variables
    --Remove the 'hh' from the directory format below to query an entire day.
    set hour_to_check = DATE_TRUNC('Hour', $pipe_altered_time)::timestamp_ntz;
    set date_directory = to_varchar($hour_to_check, '/YYYY/MM/DD/hh');
    set file_pattern = '''.*' || $date_directory || '.*''';
    select $file_pattern;
    
    -- Get the list of files in the storage account. 
    -- This query can take 10+ minutes for storage accounts with many files
    -- Note: This is currently not dynamic, and doesn't like variables. You need to manually enter the value from $file_pattern variable above
    LIST @<Table_Name>_STAGE pattern= <'.*/2020/04/23/17.*'>;
    
    -- Comparison stuff
    -- Uses the output of the previous query and compares it to what's been loaded in the table
    SELECT s.$1 as azurefilename, t.file_name as tablefilename
    FROM
    (
        table(RESULT_SCAN(LAST_QUERY_ID()))
        --If you need to specify the query id for the LIST results manually, use this instead
        --table(RESULT_SCAN('<abcabcab-00bd-74d1-0000-1d990c3c72aa>'))
    ) s
    FULL OUTER JOIN 
    (
        select distinct file_name from OUTGOING_RAW_REQUEST
        where file_name like '%' || $date_directory || '%'
    ) t
    ON charindex(t.file_name, s.$1) > 0;
    
    -- Check for mismatched files
    -- Ideally, this list should be empty
    select *
        from 
        (
            table(RESULT_SCAN(LAST_QUERY_ID()))
            --If you need to specify the query id for the comparison results manually, use this instead
            --table(RESULT_SCAN('<abcabcab-00bd-74d1-0000-1d990c3c72aa>'))
        )
        where $1 is null or $2 is null;
    
    -- If the above list is empty. You're done.
    -- If the above list returns files that were not loaded, then copy the missing files into table.
    -- Get the copy into statement. From the pipe creation statement above or using `select get_ddl('pipe', '<pipe_name>')`
    --      You'll need to cut out the pipe syntax and just use the `COPY INTO` statement
    -- the file name format should generally look like `/2020/04/23/17/20/53/28.avro` and you may need to trim the results from above to match.
    <COPY INTO OUTGOING_RAW_REQUEST FROM (SELECT PARSE_JSON(HEX_DECODE_STRING($1:Body)), '1'::numeric, metadata$filename FROM @OUTGOING_RAW_REQUEST_STAGE)>
    files = (
    '</2020/04/23/17/20/53/28.avro>',
    '</2020/04/23/17/20/52/20.avro>');