spade

NOTE: This project is no longer being updated publicly.

Spade Processor

Overview

The Spade Processor provides a service to transform and collate stat events into a format that is consistent with the current storage schema for a particular event (if it exists). If you are interested in how we create and maintain schemas take a look at twitchscience/blueprint.

The above has a lot of big words that are kind of confusing so lets start off with an example.

What the Processor Does

The processor can be described by the following diagram:

Downloads logs from S3
   +–+
   |
+––+––+     +–––––––––––––––––––––+––––+–––+
|     |     |                     |    |   ++
|     |     |                     |    +––––+       XXXXXXXXX
|     |     |                     |    |   ++  XXXXXX       XX
|     |     | Parsing Pool        |    +––––+ XX            XX
|     |     |                     |    |   ++  X            XX
|     +–––––+                     |    +––––––––+X   S3      X
|     |     |                     |    |   ++    XX           X
|     |  +  |                     |    +––––+  XX          XXXX
+–––––+  |  +–––––––––––––––––––––+––––+   ++  XXXXXXXXXXXXX
         |
         |                        +––––––––+
         | chan []byte            Writer Controller and Writers
         +

By this model if we attached a thread that simple output

222.22.222.222 [1395707641.000] data=eyJwcm9wZXJ0a...GNoZWQifQ== 09fff3e1-49eff880-535707f3-20f9e114d784a3fa

to the parsing pool, the parsing pool would receive this line and extract the data contained in the (truncated) base64 encoded json blob. This Json blob would then be transformed into a tab delimited representation of the schema for the event as described in the daemon's table_config.json file (an example of how that config looks is contained in the config directory).

This in turn will trigger an ingest of the data into Redshift.

Testing

If you are on a mac, to run the tests you need to brew install pkg-config and gzrt. If you are running this on a mac with Xcode 8.3 and Go < 1.8.1, then you need to provide -ldflags -s to your run.

Replay mode

It is also possible to replay data from an S3 bucket of deglobbed inputs which failed to process properly for some reason (errors in the Blueprint schema or bugs in Spade, for example). Once this is done, the replay script will automatically reingest the reprocessed data into Redshift, replacing the erroneous records if any. Replay can also publish to Kinesis streams, but consumers must be aware that replays are possible.

To run in replay mode:

1. Deploy a Spark cluster where each worker node has Spade installed. Spade should be isolated by only allowing one task at a time on each worker node.
2. Inside the Spade directory, run

./replay.sh MASTER_IP REDSHIFT_TARGET START END [TABLE ... | --all-tables]

where * MASTER_IP is the central node of the Spark cluster; * REDSHIFT_TARGET is the destination Redshift cluster; * START and END specify the time window to be replaced, in Pacific time in the format %Y-%m-%d %H:%M:%S; * TABLE ... specifies the tables whose records should be replaced. 3. If the results of the previous step are unsatisfactory, make necessary changes and repeat.

Replay is a 2 step process - transforming edge logs on Spark, and loading the transformed data into Redshift. For flexibility, there are 2 flags that lets you skip either steps. --skip-ace-upload will have replay do the edge log transformation, and will place the transformed files in S3, but will stop short of actually uploading into Redshift. --skip-transform will have replay skip the edge log transformation step, and directly load a runtag's transformed data into Redshift - if this flag is specified, --runtag must also be specified. For example, --skip-ace-upload is used by Tahoe replay because the transformed files are Tahoe-bound instead of Ace-bound.

For extra reliability, instead of running replay.sh locally, start a tmux session on the master node and submit from there.

We expect this tool to be used only every few months; if your needs are similar and the Spark cluster you bring up is only for this tool, consider tearing it down.

Utilities

libexec/spade_parse --config <file>

runs a mini spade server that will read spade requests at localhost:8888 and output how it parsed them.

License

see License