# Erasure code sizing guide

## Toy Setups

Capacity constrained environments, MinIO will work but not recommended for production.

| servers | drives (per node) | stripe_size | parity chosen (default) | tolerance for reads (servers) | tolerance for writes (servers) |
|--------:|------------------:|------------:|------------------------:|------------------------------:|-------------------------------:|

        "description": "MinIO instance 127.0.0.1:9000 of job minio-job has tolerance <=0 for more than 5 minutes.",
        "summary": "Instance 127.0.0.1:9000 unable to tolerate node failures"
      },
      "startsAt": "2023-11-18T06:20:09.456Z",
      "endsAt": "0001-01-01T00:00:00Z",
      "generatorURL": "http://fedora-minio:9090/graph?g0.expr=minio_cluster_health_erasure_set_tolerance+%3C%3D+0&g0.tab=1",
      "fingerprint": "2255608b0da28ca3"
    }

 * [CWAC-NetSecurity](https://github.com/commonsguy/cwac-netsecurity): Simplifying Secure Internet Access.
 * [Failsafe](https://failsafe.dev/okhttp/): Fault tolerance and resilience patterns.
 * [Flipper](https://fbflipper.com/): A desktop debugging platform for mobile developers.
 * [Fresco](https://github.com/facebook/fresco): An Android library for managing images and the memory they use.

- To be able to reliably use S3 multipart APIs feature of the SDKs without re-inventing the wheel of pre-signing the each URL in multipart API. This is very tedious to implement with all the scenarios of fault tolerance that's already implemented by the client SDK. The general client SDKs don't support multipart with presigned URLs.

- We limited the number of drives to 16 for erasure set because, erasure code shards more than 16 can become chatty and do not have any performance advantages. Additionally since 16 drive erasure set gives you tolerance of 8 drives per object by default which is plenty in any practical scenario.

import org.junit.jupiter.api.Test

/**
 * Integration test to confirm that [TaskRunner] works with a real backend. Business logic is all
 * exercised by [TaskRunnerTest].
 *
 * This test is doing real sleeping with tolerances of 250 ms. Hopefully that's enough for even the
 * busiest of CI servers.
 */
@Tag("Slowish")
class TaskRunnerRealBackendTest {
  private val log = LinkedBlockingDeque<String>()

     *
     * TODO(cpovirk): decide what the best long-term action here is: force users
     * to suppress (as we do now), stop testing entrySet().add() at all, make
     * entrySet().add() tests tolerant of either behavior, introduce a map
     * feature for entrySet() that supports add(), or something else
     */
    return Arrays.asList(
        getAddUnsupportedNotPresentMethod(),

     *
     * TODO(cpovirk): decide what the best long-term action here is: force users
     * to suppress (as we do now), stop testing entrySet().add() at all, make
     * entrySet().add() tests tolerant of either behavior, introduce a map
     * feature for entrySet() that supports add(), or something else
     */
    return Arrays.asList(
        getAddUnsupportedNotPresentMethod(),

  }

  /*
   * Our implementation works fine with a null `dir`. However, there's nothing in the documentation
   * of the supertype that suggests that implementations are expected to tolerate null. That said, I
   * see calls in Google code that pass a null `dir` to a FilenameFilter.... So let's declare the
   * parameter as non-nullable (since passing null to a FilenameFilter is unsafe in general), but if

      immediate_execution_ctx_.reset(ctx_raw);
    }

    // Computing numerical gradients with TensorFloat-32 is numerically
    // unstable. Some forward pass tests also fail with TensorFloat-32 due to
    // low tolerances
    enable_tensor_float_32_execution(false);
  }

  AbstractContextPtr immediate_execution_ctx_;
  GradientRegistry registry_;
  Status status_;

 public: